Americas Cup Optimisation

S TOCHASTIC OPTIMISATION OF
A MERICA’ S C UP C LASS YACHTS

BY
Andrew Phillip Mason
A THESIS SUBMITTED IN FULFILMENT OF THE REQUIREMENTS FOR
THE DEGREE OF
DOCTOR OF PHILOSOPHY
AUSTRALIAN MARITIME COLLEGE
UNIVERSITY OF TASMANIA
February 2010
COPYRIGHT © 2010
ANDREW MASON
ALL RIGHTS RESERVED
Stochastic Optimisation of America’s Cup Class Yachts
Declaration of originality
This thesis contains no material which has been accepted for a degree or diploma by
the Australian Maritime College, University of Tasmania or any other institution,
except by way of background information and duly acknowledged in this thesis, and
to the best of Andrew Mason’s knowledge and belief, no material previously
published or written by another person, except where due acknowledgement is
made in the text of this Thesis.
Much of the experimental data reproduced within this thesis is the property of the
Alinghi America’s Cup team and is reproduced with permission.
_________________________ February 12, 2010
·i·
Statement of authority of access

Printed copies of this thesis are not to be made available for viewing, loan or
copying prior to September 1, 2012. Subsequent to that period, the thesis may be
made available for loan and limited copying in accordance with the Copyright Act
1968.
Electronic copies of this thesis are not to be released, published or distributed in any
form prior to September 1, 2012.
· ii ·
Abstract
This thesis describes the design and implementation of an optimisation system for
America’s Cup Class (ACC) yachts. The system, named VESPA, uses a measure of
merit that closely approximates the actual America’s Cup race format; a round-robin
match-racing tournament, held over many races between a population of candidate
designs, using a stochastic wind model. VESPA was used by the Alinghi team to
provide design recommendations for the 2007 America’s Cup.
The optimisation of racing yachts is a problem that has been considered resistant to
full analysis due to its complexity. Consequently, attempts at yacht design
optimisation to date have been restricted to simplified subsets of the problem.
While Velocity Prediction Programs (VPP) have been widely used to provide details
of sailing performance for one or more yachts, the statistical models on which these
programs are based have not been sufficiently accurate to allow optimisation of hull
shapes. Other efforts to automate yacht design optimisation have used an objective
function that evaluates the performance of each boat using Computational Fluid
Dynamic (CFD) analysis of the hull. This approach suffers from long execution
times, which may result in the adoption of a restricted measure of merit, such as
hull resistance at a small number of forward speeds, heel and yaw angles.
In order to permit the use of the chosen measure of merit while retaining acceptable
performance, a sparse sample of designs, derived from a parent hull using a novel
parametric transformation method, had their hydrodynamic characteristics
calculated by the SPLASH potential flow code. The output from SPLASH was
subsequently used to train a set of neural-network based hydrodynamic metamodels
for use by the VPP. The need to assess a population of designs for the tournament-
based measure-of-merit makes the problem well suited to stochastic, population
based optimisation methods. As a result, a Genetic Algorithm (GA) was chosen to
perform the optimisation, using a parsimonious Race Modelling Program (RMP) to
simulate a tournament of races based on performance data provided for each boat
by the VPP.
Each component within the VESPA system was validated to ensure confidence in
the optimisation results. Optimisation runs were performed over several months
using multiple parent models to investigate the effect of changes to various design
variables. Finally, a design optimised by VESPA was tank tested at 1/3 scale,
confirming the improvements over its parent design predicted by SPLASH.
VESPA proved itself capable of making genuine design improvements to an existing

parent model while retaining reasonable execution times. VESPA also revealed
several unexpected insights into the nature of the solution space for the design of
ACC yachts, including multiple optima and the potential for intransitivity in the
solution when interactions between boats at rounding marks are considered.
· iii ·
Acknowledgements
I would like to express my heartfelt gratitude to my wife Mandy, who despite her
opinion that Ph.D. stood for “Phile for Divorce”, showed immense patience in
allowing me the time necessary to complete this work.
My thanks to my mother, for her patience and gentle guidance over the years, and
to my late father, for buying me a boat and teaching me to sail as a child, instilling
in me an enduring love of the water.
I would also like to thank my supervisor, Dr. Giles Thomas, for the disciplined
approach he brought to the task and for his quiet encouragement, without which
this thesis may have foundered.
Thanks also to Barbara Backhouse for her meticulous proofreading skills.
Finally, I would like to thank the Alinghi design team, in particular Manuel Ruiz de
Elvira, Grant Simmer, Rolf Vrolijk and Michael Richelsen, for the support and
assistance they provided, without which this work would not have been possible.
· iv ·
Foreword
“Every point of sailing suggests an appropriate and different form of hull. The
shape that is well adapted for one kind of weather is ill adapted for another sort;
vessels that move as by magic in light airs may be of little use in a whole sail
breeze; one that is by no means a flier in smooth water may be very hard to beat
in a seaway.
In short, a vessel must be light enough to be driven easily by a moderate breeze,

stiff enough to stand up to her canvas in a hard wind, shallow enough to be
docked with ease and to run with speed. She must have depth enough to hold her
up to windward, breadth enough to give her stability; she should be long enough
to reach well, and short enough to turn well to windward; low in the water so as
not to hold too much wind, with plenty of freeboard to keep the sea off her decks.
The satisfaction of any one requirement necessitates something antagonistic to

some other requirement equally clamorous for satisfaction. Your vessel, to be perfect,
must be light, of small displacement, and with the centre of gravity brought very
low; she must also have large displacement, and the ballast must not be too low, in
order that she may be easy in a seaway; she must be broad, narrow, long, short,
deep, shallow, tender, stiff. She must be self-contradictory in every part. A sailing
yacht is a bundle of compromises, and the cleverest constructor is he who, out of a
mass of hostile parts, succeeds in creating the most harmonious whole.
It is not strange that designers pass sleepless nights, and that anything like finality
and perfection of type is impossible to conceive. No wonder that yacht designing is
a pursuit of absorbing interest.”
Lord Dunraven, challenger for the 1893 and 1895 America’s Cups.
·v·
Contents
1 Overview 1
1.1 PROBLEM DEFINITION 3
1.2 BACKGROUND AND CONTEXT 4
1.3 OPTIMISATION – A DEFINITION 7
1.4 DESIGN OPTIMISATION – STATE OF THE ART 8
1.4.1 Early ship design optimisation 8
1.4.2 Analysis methods 9
1.4.3 Optimisation methods 14
1.4.4 Hull shape variation methods 20
1.4.5 Sailing-yacht design optimisation 25
1.5 DESIGN OPTIMISATION SUMMARY 26
1.6 OUTLINE OF THESIS 27
1.7 CONTRIBUTION 28
2 Measure of Merit 29
2.1 MEASURE OF MERIT OR OBJECTIVE FUNCTION? 30
2.2 UPRIGHT RESISTANCE 30
2.3 HEELED LIFT AND DRAG 31
2.4 VELOCITY PREDICTION PROGRAMS 32
2.5 COMPARISON PLOTS 34
2.6 RACE MODELLING PROGRAMS 35
2.7 MONTE CARLO RACE MODEL SIMULATIONS 36
2.8 TOURNAMENT MODELLING 38
2.9 WEATHER MODEL 39
2.10 MEASURE OF MERIT - SUMMARY 39
3 Constraints 40
3.1 THE AMERICA’S CUP CLASS RULE 40
4 Data Approximation 48
4.1 DATA CHARACTERISTICS 49
4.2 METAMODELS 50
4.2.1 Metamodel candidates 53
4.2.2 Metamodelling – practical usage 58
4.2.3 Metamodel selection 58
4.3 DESIGN OF EXPERIMENTS 59
4.3.1 Quasi-random methods 60
4.4 HULL SHAPE REPRESENTATION 62
4.5 HULL SHAPE VARIATION 65
4.5.1 Direct control point manipulation 65
4.5.2 Blending 66
4.5.3 Parametric modelling 68
· vi ·
4.5.4 Parametric transformation 71

4.5.5 VESPA hull shape transformation 77
4.6 SUMMARY 80
5 Performance Evaluation 81
5.1 VPP 82
5.2 VESPA RACE MODELLING PROGRAM 84
5.2.1 Match racing tactics 84
5.2.2 RMP background 89
5.2.3 Proposed VESPA RMP methodology 92
5.2.4 Accounting for uncertainty 99
5.2.5 Summary 106
6 Optimisation Algorithm 107
6.1 OPTIMISATION METHODS 107
6.2 GRADIENT AND PATTERN SEARCH METHODS 108
6.3 STOCHASTIC METHODS 109
6.3.1 Simulated Annealing 109
6.3.2 Particle Swarm Optimisation 110
6.3.3 Evolutionary Algorithms 111
6.4 PROPOSED GA METHODOLOGY 113
6.4.1 GA parameters 113
6.4.2 Recombination 114
6.4.3 Mutation 116
6.4.4 Selection 117
6.4.5 Population size 117
6.4.6 Fitness sharing 118
6.4.7 Elitism 118
7 Implementation 122
7.1 INTEGRATION OF COMPONENTS 123
7.2 PARAMETRIC TRANSFORMATION FUNCTION 124
7.2.1 Hierarchical Free-Form Deformation 125
7.3 SYSTEMATIC SERIES CREATION 128
7.4 SPLASH ANALYSIS 130
7.5 NEURAL NETWORK METAMODEL CREATION 131
7.6 VPP 132
7.6.1 DLL link from VPP to optimiser 133
7.7 RACE MODELLING PROGRAM 133
7.7.1 Wind model 135
7.7.2 Wallying 135
7.8 GENETIC ALGORITHM 136
7.9 USER INTERFACE 137
· vii ·
8 Verification & Validation 139

8.1 PARAMETRIC TRANSFORMATION VALIDATION 140
8.2 SPLASH VALIDATION 140
8.3 METAMODEL TRAINING & VALIDATION 141
8.4 NEURAL NETWORK VALIDATION 146
8.5 NEURAL NETWORK DLL VALIDATION 146
8.6 GA VALIDATION 147
8.7 VPP VALIDATION 149
8.8 RMP VALIDATION 150
8.9 VESPA SYSTEM TESTING 151
8.9.1 Objectives and methodology 151
8.9.2 Optimisation runs 152
8.9.3 Results and validation 153
8.9.4 Validation model conclusions 158
9 Results & Discussion 159
9.1 CONSEQUENCES OF TRAILING BOAT PENALTIES 160
9.1.1 Genetic algorithm modification 165
9.2 PARENT MODEL 1, MAY 28, 2006 167
9.2.1 Trailing boat penalties 171
9.3 PARENT MODEL 2, JULY 30, 2006 173
9.4 PARENT MODEL 3, AUGUST 18, 2006 177
9.5 ANALYSIS OF RESULTS 178
9.5.1 Sensitivity of design factors 178
9.5.2 Choice of parameters for tank testing 180
9.5.3 Validation against tank test data 181
9.5.4 Outcome 183
9.6 SUMMARY 185
10  Conclusions 186
10.1 CRITICAL ASSESSMENT & CONTRIBUTION 188
10.2 ADVANTAGES AND DISADVANTAGES OF VESPA 189
10.3 FUTURE WORK AND OUTLOOK 190
Publications 192
References 193
Appendix 1 – XML Metamodel File Structure 210
Appendix 2 – Neural Network Class Structure 215
Appendix 3 – VPP DLL Interface 219
Appendix 4 – Race Time Calculation 221
Appendix 5 – Tournament Modelling 222
Appendix 6 – Genetic Algorithm Classes 225
· viii ·
Figures
Page
Figure 1. Design Spiral 4
Figure 2. Alternating variable method 5
Figure 3. Alternating variable method with correlated variables 6
Figure 4. SHIPFLOW zones, from Janson and Larsson (1996) 11
Figure 5. Superiority of GA versus SQP, from Hirata (2004) 16
Figure 6. Illustration of Pareto optimality, from Zitzler (2002) 17
Figure 7. Longitudinal undulations in hulls optimised for a single speed,
from Chen (2004) 18
Figure 8. Optimisation of a Wigley hull at different Froude numbers,
from Percival et al. (2001) 19
Figure 9. Resistance curves for hulls optimised by Percival et al. (2001) 19
Figure 10. Barycentric blending of hull shapes, from Neu et al. (2000) 21
Figure 11. An FFD used for deformation of an aircraft wing (© INRIA Opale
Project-Team 2007, reproduced with permission) 23
Figure 12. An FFD lattice enclosing the bulbous bow of a cruise ship,
from Peri and Campana (2005a) 23
Figure 13. Non-linear deformation of primate skulls, from Thompson (1917) 24
Figure 14. Variation of fish forms using linear and non-linear transformations,
from Thompson (1917) 24
Figure 15. Optimal hull shapes for different heel angles,
from Baik and Gonella (2005) 31
Figure 16. Polar performance curves, © ORC 2004, reproduced with permission 32
Figure 17. Pareto optimal ACC designs, from Jaquin et al. (2002) 33
Figure 18. Performance comparison plot 34
Figure 19. ACC Hull Measurements 42
Figure 20. Rated Length (L) versus Length Between Girths (LBG) 43
Figure 21. Rated Sail Area (S) versus Measured Sail Area (SM) 43
Figure 22. Weight (W) as a function of Length Between Girths (LBG) and
Measured Sail Area (SM) 44
Figure 23. Usable ranges for Length Between Girths (LBG) and Measured Sail
Area (SM) 44
Figure 24. ITA-94, an example of ACC hull design extremes 46
Figure 25. ACC rig configuration (© Ivo Rovira 2007, reproduced with
permission) 46
Figure 26. Objective function contaminated by noise, from Keane and Nair (2005) 52
Figure 27. Neuron with sigmoidal activation function 55
Figure 28. Multi-layer perceptron architecture 56
Figure 29. Radial Basis Function network architecture 57
· ix ·
Figure 30. Full factorial, Box-Benkhen and central composite experimental designs 59
Figure 31. Quasi-random sampling methods 60
Figure 32. NURBS surface and corresponding control point net 63
Figure 33. Blending of ACC bow sections 67
Figure 34. An ACC canoe body optimised using FRIENDSHIP,
from Harries et al. (2001) 69
Figure 35. Comparison of parametrically modelled ACC hull with typical 2007
ACC design 69
Figure 36. Cross sectional shapes of optimal hulls, from Baik and Gonella (2005) 70
Figure 37. FFD lattice enclosing a tessellated sphere 73
Figure 38. Deformation of an FFD lattice 74
Figure 39. Free Form Deformation of an embedded curve 75
Figure 40. Effects of FFD weight and control point changes on longitudinal
volume distribution 76
Figure 41. Deformation of hull section using a global FFD 78
Figure 42. Transformation of an ACC hull using a hierarchical FFD 79
Figure 43. Forces acting on a sailing yacht 82
Figure 44. Wind shadow effect of leading boat, A 84
Figure 45. Safe leeward position 85
Figure 46. A tacking into a tight cover on B 86
Figure 47. Tactical control of A by port tack yacht, B 86
Figure 48. Trailing boat options at the layline 87
Figure 49. Leeward gate options for trailing boat 88
Figure 50. Right-of-way advantage at the start 88
Figure 51. Win probabilities for various Stars & Stripes Measurement Waterline
Lengths (MWL) against opponents of various lengths (January Perth
conditions), from Oliver et al. (1987) 90
Figure 52. Windward mark rounding, America's Cup 2007, race 5, leg 1 96
Figure 53. Windward leg details, America’s Cup 2007, races 1 to 3 98
Figure 54. Polar performance curve with VMG 104
Figure 55. Polar performance curve for a shifted wind direction, showing
advantage of wallying 104
Figure 56. A taxonomy of stochastic optimisation methods 109
Figure 57. Single-point and dual-point crossover, from Morgan (1916) 114
Figure 58. Binary recombination and mutation 114
Figure 59. VESPA logic flow 123
Figure 60. Control points mapped into the unit square 125
Figure 61. Bounding boxes displayed for each control point column in C 126
Figure 62. FFD planes enclosed by FFD lattice P 126
·x·
Figure 63. Deformed FFD lattice P’ with updated weights ′, resulting in

deformations to the embedded planar NURBS surfaces Q’ 127
Figure 64. Deformed hull surface defined by updated control point network C' 127
Figure 65. A simple systematic series of hulls, varying in CP, BWL, CM , T and flare 128
Figure 66. Hull-series generation spreadsheet 129
Figure 67. Comparison of wind velocity distributions 135
Figure 68. VESPA application 137
Figure 69. VESPA optimisation controls 138
Figure 70. VESPA results window 138
Figure 71. Extended neural network training 143
Figure 72. Correlation plot of neural net prediction versus actual, showing outliers 144
Figure 73. Error values for different numbers of GA generations 148
Figure 74. Average Relative Error values for variations in population size 149
Figure 75. Family of 25 hull design variations 151
Figure 76. Performance comparison plots of test models versus validation model 155
Figure 77. Validation spreadsheet controls 156
Figure 78. Metamodel validation graphs 157
Figure 79. Drag errors for 11 knots velocity, 1.5 degrees of yaw 158
Figure 80. Investigation into the effects of Trailing Boat Penalties (TBP) 161
Figure 81. Optimal beam for a simple yacht performance model 162
Figure 82. Effect of trailing boat penalties on race performance 163
Figure 83. Win/loss ratio for a range of BWL values, utilising trailing boat penalties 164
Figure 84. Effect of trailing boat penalties 165
Figure 85. Investigation into the effects of exemplars and/or Hall of Fame 166
Figure 86. Test hulls R1 (top) and R2 (bottom) 169
Figure 87. Performance comparison plots of test models versus parent model 1 170
Figure 88. High CM designs based on parent model 1 171
Figure 89. BWL sensitivity analysis 174
Figure 90. CM sensitivity analysis 175
Figure 91. Comparison plots of X1 and X3 versus parent model 2 176
Figure 92. CM sensitivity analysis 179
Figure 93. Flare sensitivity analysis 179
Figure 94. BWL sensitivity analysis 180
Figure 95. Tank model MT 181
Figure 96. MS parent model versus MT, SPLASH and tank results comparison 182
Figure 97. MO versus transformed design MOX3 184
· xi ·
Tables
Page
Table 1. Early Optimisation Studies in Basic Ship Design 9
Table 2. Summary of Developments in Yacht Design Optimisation 26
Table 3. Recommendations For Model Choice And Use, After Simpson (2001) 53
Table 4. Parameters Used to Describe an ACC Hull, From Harries et al. (2001) 68
Table 5. Tacks Performed During Races 1-7, America's Cup 2007 95
Table 6. Times for Races 2 and 3, America's Cup 2007 97
Table 7. Results of Two-Boat Testing, May-June 2006 102
Table 8. Comparison of Neural Network Training Algorithms 142
Table 9. Convergence Properties of LM and QN Methods 142
Table 10. Neural Network Prediction Errors – Parent Model 3 146
Table 11. Genetic Algorithm Performance for Synthetic VPP Data 147
Table 12. Population Size Study Outcomes 148
Table 13. Change in Optimal BWL with the Incorporation of Random Noise 150
Table 14. Parameter Variation Limits – Validation Model 151
Table 15. Results Summary – Validation Model 153
Table 16. Test Design Parameters – Validation Model 154
Table 17. Parameter Variation Limits – Parent Model 1 167
Table 18. Parent Model 1 Results 167
Table 19. Test Model Parameters – Parent Model 1 168
Table 20. Results With Trailing Boat Penalties Included – Parent Model 1 171
Table 23. Test Hull Parameter Values – Parent Model 2 175
Table 26. Parent Model 3 Tests Versus Exemplars, Mean Wind Velocity = 11
knots, Event SD = 4 knots and Race SD = 1.4 knots 177
Table 27. Effects of Wallying on design optima – Parent Model 3 178
Table 28. Tank Model Candidate Parameter Values – Parent Model 3 180
· xii ·
Nomenclature – naval architecture
AW waterplane area
BWL maximum beam of waterline
CF coefficient of frictional resistance [ITTC-57 correlation line]
CIL longitudinal waterplane inertia coefficient
CM midship area coefficient
CP prismatic coefficient
CR coefficient of residuary resistance
CT coefficient of total resistance
CW coefficient of wave resistance
CWP waterplane area coefficient
Fr Froude number, /√
Frd depth Froude number, /√
FB distance from the forward perpendicular to centre of buoyancy

g acceleration due to gravity
LCB non-dimensional longitudinal centre of buoyancy, CB / WL
LCF longitudinal centre of flotation

LCG longitudinal centre of gravity
LWL waterline length
L/∇1/3 slenderness ratio
Re Reynolds number [ / ]
RRU residuary resistance upright
RRϕ residuary resistance heeled
RV viscous resistance
S wetted surface area
TC draft of canoe body
VB boat velocity
∇ volume of displacement
fluid kinematic viscosity
ϕ heel angle
ρ fluid density
· xiii ·
Nomenclature – yacht design
βTW true wind angle

βAW apparent wind angle
VAW apparent wind velocity
VMC velocity made good relative to the direction of the course
VMG velocity made good relative to the direction of the wind
VTW true wind velocity
λ leeway angle
Nomenclature - statistical
SD or Standard Deviation ∑ ( ) , where ∑
AAE Average Absolute Error (∑ | |) /
ARE Average Relative Error (∑ | |

)/
Acronyms – design & analysis

ACC America’s Cup Class
CAD Computer Aided Design
CFD Computational Fluid Dynamics
DOE Design of Experiments
FFD Free Form Deformation
NURBS Non-Uniform Rational B-Splines
PDF Probability Density Function
RANS Reynolds Averaged Navier-Stokes
RAO Response Amplitude Operator
RMP Race Modelling Program
VESPA Virtual Evolution based Sailing Performance Analysis
VPP Velocity Prediction Program
· xiv ·
Acronyms - optimisation methods

BFGS Broyden Fletcher Gordfarb Shanno
DFP Davidson Fletcher Powell
EA Evolutionary Algorithm
ES Evolution Strategies
GA Genetic Algorithm
GP Genetic Programming
LM Levenberg Marquardt
MMA Method of Moving Asymptotes
MOGA Multi-Objective Genetic Algorithm
NLP Non-Linear Programming
PSO Particle Swarm Optimisation
QN Quasi Newton
SA Simulated Annealing
SQP Sequential Quadratic Programming
SUMT Sequential Unconstrained Minimisation Technique
Acronyms - approximation methods

ANN Artificial Neural Networks
MARS Multivariate Adaptive Regression Splines
PR Polynomial Regression
RBF Radial Basis Functions
RSM Response Surface Methods
SVM Support Vector Machines
· xv ·
America’s Cup Class - rule variables
ABC aft beam correction

AG aft girth
AGC aft girth correction
AGS aft girth station
ALM aft length mark (positioned at the aft end of LBG)
DSP ACC displacement in cubic metres
E1..E5 mainsail girths
FBC forward beam correction
FG forward girth
FGC forward girth correction
FGS forward girth station
FLM forward length mark (positioned at the forward end of LBG)
FP forward penalty
G girth component of LM
I height of foretriangle
J base of foretriangle
L rated length
LBG length between girth stations
LM measured length
MWL measurement water line
P mainsail luff length
S rated sail area
SM measured sail area
W weight of yacht
Φ mean of the angles port and starboard of the topsides measured at AGS,
relative to the vertical, in degrees
Θ mean of the angles port and starboard of the topsides measured at FGS,
relative to the vertical, in degrees
· xvi ·
Chapter 1 · Overview
“The America’s Cup is a design race. At this level, against this calibre of sailors, if
you have a slightly faster boat you are going to beat them. At every America’s Cup
I’ve done, the fastest boat has won.”
Brad Butterworth, Alinghi skipper and three times America’s Cup winner.
1  Overview
The 2007 America’s Cup involved 12 teams whose budgets totalled more than
$US1 billion. Of this figure, approximately 25% was expended on design, research
and development aimed at winning the America’s Cup final, a series of 7 yacht races
held over a period of 2 weeks.
Despite this large investment of time and money, design optimisation of ACC
yachts has remained primarily a manual process. This is not to say that the design of
ACC yachts does not involve sophisticated technology; ACC design teams use some
of the most powerful computational fluid dynamics and structural analysis software
available, and utilise clusters of high performance computers to perform a volume of
calculations that would have been unthinkable 20 years ago.
Rather, the problem is that there has been little work done on automating the
exploration of the design space. Despite the emphasis on computational tools, the
overall approach to design optimisation has changed little since computer based
simulation tools were adopted.
Prior to the 1980s, the design development of America’s Cup yachts primarily relied
on tank testing, with only limited computing power available for numerical analysis.
With computer costs plummeting and computing power increasing dramatically in
the early eighties, it became possible to consider simulating, on a computer, the
experiments previously performed in a towing tank.
·1·
Prior to the introduction of Computational Fluid Dynamics (CFD) software, yacht

design optimisation had been based on a matrix of proposed designs that were built
as scale models and tank tested. This allowed some parameter variation to take place
and conclusions to be drawn from the result. However, the time and cost of tank
testing prohibited the systematic optimisation of a hull design.
The quality of the finished hull design was very much dependent on the skill and
experience of the designer, and the conclusions that could be drawn from a small
quantity of data and a large amount of intuition. The insight of the yacht designer
was paramount and it is no surprise that the post-war years of the America’s Cup
were dominated by the designs of one exceptional designer, Olin Stephens, who was
involved in the design of all but one America’s Cup winner from 1937 to 1980.
One of the first America’s Cup yachts to gain significant advantage from numerical
simulations was Australia II, winner of the 1983 America’s Cup, with the intuition
of her designer Ben Lexcen complemented by the computational work of two
Dutch engineers, Peter van Oossanen and Joop Slooff, (van Oossanen 1985). Since
then, cup teams have devoted ever-increasing amounts of effort to both
hydrodynamic and aerodynamic simulations of the hull and rig designs.
Although much work has been done in the past 25 years with CFD analysis of yacht
designs, surprisingly little has been done to automate the process to allow the
computer to search for an optimal design for a given set of weather conditions. To
some degree this has been due to the large amount of computing power required. It
is also a consequence of the enormous complexity of the yacht design and analysis
process. To solve the problem effectively there is a need to reduce this complexity,
where appropriate, in order to perform the optimisation in a reasonable amount of
time on available computer hardware.
If a workable design optimisation system could be developed, several benefits would

result. Firstly, optimisation of new designs would occur more rapidly. It has taken
more than fifteen years for America’s Cup Class designs to evolve from wide, heavily
flared yachts with soft bilges, to narrow, hard-bilged hulls with vertical topsides. It is
possible that the optimality of this area of the design space could have been rapidly
discovered by a well-designed optimisation system.
Once a design is close to optimal, it becomes more difficult to discover design

variations that will result in improvements in performance. In this area of
diminishing returns, an automated optimisation system may also have significant
benefits, allowing design improvements to be achieved without the expenditure of
an excessive number of man hours.
In addition, if the system is able to perform global optimisation, thoroughly

exploring the available design space, novel design approaches may be uncovered that
might not be obvious to a designer employing traditional manual methods.
·2·
1.1 Problem definition

This thesis describes the research, design and development of a design optimisation
system for America’s Cup Class (ACC) yachts. The system, named VESPA (short for
Virtual Evolution based Sailing Performance Analysis), has been developed in
collaboration with the Alinghi America’s Cup team. VESPA was used to make design
recommendations for Alinghi’s successful defence of the 2007 America’s Cup.
VESPA was developed with the aim of assisting a yacht designer to determine the
optimal design parameters required for a yacht to succeed in winning the America’s
Cup. In this regard, VESPA does not attempt ab initio design of the hull shape.
Rather, VESPA attempts to refine a specific parent hull by varying a small number of
key parameters, while retaining key design features that may have resulted from the
expenditure of thousands of hours of manual refinement and research.
The first step in such an optimisation problem is determining how to ascertain

whether one yacht is superior to another; that is, what measure of merit should be
used? As pointed out by Harries (2001a):
“For an ACC yacht the design evaluation becomes a challenging task in itself since
a probabilistic measure of merit ought to be considered... however, an
extraordinary amount of computation will be required to determine this ultimate
measure of merit and to optimise for it.” (Harries et al. 2001a, pp.11-12)
Harries highlights the two key issues in creating a system such as VESPA; the correct
choice of measure of merit and the problem of system performance. Using modern
Reynolds Averaged Navier-Stokes (RANS) CFD codes directly in an optimisation
loop could conceivably result in execution times that stretch into weeks or months
for the fastest optimisation methods. Korpus (2004) expressed similar concerns
regarding measures of merit and the performance of RANS codes:
“While faster computers have made RANS analyses possible, most applications to
date fall short of being practical. If an America’s Cup designer is to improve boat
speed, he or she must analyze hundreds of design alternatives - not the few isolated
samples usually associated with RANS. And even when a large number of flow
analyses are available, the measures of merit required to rank designs are not
obvious RANS outputs like flow detail or drive force.” (Korpus 2004, p.249)
Unfortunately, the most appropriate measure of merit for this particular problem,
the win/loss probabilities for an entire fleet of boats competing in a round-robin
match-racing tournament, requires some of the slowest optimisation methods. This
is due to it being necessary to maintain a population of designs, rather than
attempting to optimise a single design. This combination of slow analysis with slow
optimisation methods makes the option of direct CFD analysis unattractive, even
when using solvers running in parallel on a cluster of computers.
·3·
1.2 Background and context

In general, the design of ships and boats can be considered an optimisation problem
in which a multitude of factors, including size, cost, speed, seaworthiness, stability,
comfort, manoeuvrability, accommodation space, crew workload and even aesthetics
are traded off against one another. The interaction between these objectives and
constraints can be complex, with some complementary and many others conflicting.
Yacht design is a specialised discipline within the field of naval architecture due to
the need to account for the presence of sails that apply both driving force and
heeling moment to the yacht. These forces affect the stability, motions and
appendage design of the vessel, the result being an extremely complex set of
constraints and objectives that need to be satisfied. For this reason, yacht design has
often been regarded as more of an art than a science, due to the number of variables
and their resultant interactions being so numerous as to defy complete analysis.
Racing yacht design adds further complexity, with a mix of performance

characteristics, both upwind and downwind, in a variety of sea conditions and wind
velocities, needing to be considered.
The design process used by naval architects to satisfy multiple, conflicting objectives
has traditionally proceeded through a series of iterations of what has been referred to
as a design spiral (Evans 1959), illustrated in Figure 1.
Figure 1. Design Spiral
Starting at the conceptual design stage, the naval architect defines a preliminary set
of hull lines and performs various forms of analysis to determine whether the design
will satisfy requirements for such objectives as stability, seakeeping, strength,
performance and construction cost. Modifications are made to the lines,
compartmentation, arrangements, equipment and structure, and the process is
·4·
repeated, each loop around the spiral resulting in a more refined design. This
process continues through the various design phases in a sequential and iterative
manner until all design criteria are adequately satisfied, concluding with the
production design phase.
Although this process has served the naval architecture community well for many
years, it is time consuming and not guaranteed to result in an optimal design.
Balancing multiple, conflicting objectives is a difficult task requiring a great deal of
skill and experience on the part of the designer, and time and cost limitations often
result in less than optimal designs being constructed.
The design spiral approach can be viewed as an extension of the widely held idea
that when seeking to improve a system that has multiple parameters, it is important
to vary only one parameter at a time while keeping the others constant, so that the
effect of any change can be easily quantified.
Within the field of optimisation, this approach is referred to as the alternating

variable method (Fletcher 1987). It typically progresses by searching a single
parameter direction until no further improvement is seen, followed by a switch to a
different parameter, which is again searched until once again no further benefit is
seen. At this point additional parameters may be searched or the process may repeat
from the first parameter until no improvement can be made.
An example of the alternating variable method incorporating two parameters is

illustrated in Figure 2. Parameter P1 is searched until a maximum is found, followed
by a search using parameter P2 resulting in a solution that is very close to the
optimum. This may be improved by repeating the cycle one or more times. On
well-behaved functions that have a reasonably small number of independent
parameters, the approach rapidly converges on the optimal solution.
Figure 2. Alternating variable method
The problem with this method is that it ignores the possibility of correlation
between the variables. This correlation effectively causes a ridge or valley to form
·5·
diagonally through the solution space. When this occurs, it causes the search in the
current search direction to destroy the property that the current point is the
minimiser or maximiser in previously used directions, (Fletcher 1987).
This leads to oscillatory behaviour of the alternating variable algorithm, illustrated

in Figure 3.
Figure 3. Alternating variable method with correlated variables
In this case, an initial search along P1 locates a ridge in the solution surface that
runs diagonally relative to the variables being searched. A switch to a search of P2
does not go far before it also reaches a maximum. A further switch back to the P1
search direction makes another small gain, followed by small gains on each
successive alternation of the search direction.
Interdependence of design variables may confound any search for an optimal design,
as a change to one variable needs to be matched by simultaneous changes to the
correlated variables. The result is an optimisation method that either progresses
slowly or fails to locate an optimum at all.
The alternating variable method is seldom used as a computer based optimisation

algorithm, and automated optimisation methods have been designed to avoid these
pitfalls. The traditional naval-architecture design spiral is functionally equivalent to
the alternating variable method and suffers from similar shortcomings when
parameters are not independent. Unfortunately, in the design of ships and boats,
interdependence of design parameters is the rule rather than the exception.
In contrast, many computer based optimisation algorithms, for example the

conjugate gradient method, (Hestenes and Stiefel 1952), attempt to determine the
best search direction for convergence. The potential for automated optimisation
procedures to find solutions where manual methods fail suggests that there may be
scope for an automated, computer based optimisation system to improve on the
best designs achieved using the conventional design spiral.
·6·
1.3 Optimisation – a definition

The term optimisation has been widely used within the field of marine design to
describe many different processes. These include manual searches of design
alternatives, the selection of the best design from a systematic series or matrix of
designs, or an automated search using computer-based simulations.
To clarify the meaning of the term as used within this thesis, optimisation is defined
as the process of finding a good (and possibly the best) feasible solution to a
problem, based on the evaluation of an objective criterion.
An optimisation problem with a single objective can be stated as:

( ),
having the following components:
Measure of merit or objective function: criteria by which a solution is assessed:

( )
Free variables, decision variables, design variables: independent variables that can
be modified directly and which uniquely describe the optimisation problem:
( , ,…, )
Constraints: define the boundaries of the feasible regions of the design space. These
can be categorised as:
• Bounds:
for i 1, 2, . . . , n
• Equality constraints:
( ) 0 for j 1, 2, . . . , m
• Inequality constraints:
( ) 0 for k 1, 2, . . . , p
Dependent variables: are those not directly controlled, but dependent on the
values of the free variables.
Parameters: additional values and conditions that are not under the direct control
of the optimisation process.
Constants: values that do not change during the optimisation process.
·7·
1.4 Design optimisation – state of the art

During the past 30 years, significant research has been performed within the field of
naval architecture on the subject of automated design optimisation of marine vessels.
Although much work has taken place in such areas as optimisation of structure,
compartmentation and propulsive systems, these domains are outside the scope of
this thesis, whose focus is on those factors that specifically influence the
hydrodynamic performance of the vessel.
In order to review the existing work in this field it is necessary to be aware of the key
areas in which the various optimisation approaches examined may differ. These can
be summarised as-
• What are we trying to improve? That is, what measure of merit should we
choose?
• How do we measure it? What analysis method should we use to calculate the
measure of merit?
• How do we improve it? What optimisation method do we select?
• How do we vary design parameters in order to make improvements? What
hull geometry representation and variation method should we adopt?
Researchers have found many different answers to these questions during the past
three decades, and research and development is ongoing. Despite dramatic advances
in computer hardware and software performance, the goal of fast, flexible and
accurate simulation and optimisation is yet to be achieved. Creating a system that
has acceptable performance and provides useful results is a significant challenge.
Before reviewing the research on the optimisation of sailing yachts in general, and
ACC yachts in particular, it is appropriate to examine the research that has been
performed on the optimisation of other forms of marine vessels. Although yacht
design is a highly specialised area within the field of naval architecture,
developments in design and analysis methods for full-sized ships share many
common features.
1.4.1 Early ship design optimisation
Prior to 1980 there are surprisingly few instances in the literature describing the use
of optimisation methods within the field of naval architecture. Much of the early
research into ship design optimisation was limited to the investigation of variations
in principal dimensions at the initial design stage, rather than the detailed
refinement of hull shapes.
·8·
The measure of merit typically used was based on ship economics, rather than
focussing on specific objectives such as resistance or ship motions. A summary based
on Nowacki (2001), showing some early optimisation studies in basic ship design, is
shown in Table 1.
Table 1. Early Optimisation Studies in Basic Ship Design
Authors Year Ship Type Criterion Approach

Murphy et al. 1965 GCS LCC Systematic variation and
interpolation
Mandel and 1966 T, GCS LCC, RFR, NPV Exponential random search,
Leopold unconstrained
Gallin 1967 GCS Several choices Systematic variation
Kuniyasa 1968 T, BC CRF Parametric study
Gilfillan 1969 BC RFR Parametric study
Puchstein 1969 GCS & Building cost, Non-linear model,
Diverse owner’s criteria implicit constraints
Nowacki et al. 1970 T RFR NLP: SUMT & DS
Fisher 1972 BC RFR NMS, implicit constraints
Lorentz 1973 T CRF, operational NLP: SUMT & DS
cost
Jagoda 1973 Diverse NPV NMS
Söding and 1974 BC AAC NLP with slack variables
Poulsen
Nowacki and 1976 T,BC, GCS RFR NLP: feasible directions &
Lessenich penalty functions
Kupras 1976 BC Building cost Parameter studies & adapted DS
Abbreviations
BC = bulk carrier AAC = average annual cost DS = direct search
GCS = general cargo ship CRF = capital recovery factor NLP = non-linear programming
T = tanker LCC = life cycle cost NMS = Nelder-Mead simplex
NPV = net present value SUMT = sequential
RFR = required freight rate unconstrained minimisation
technique
Parsons (1975) was among the first to publish a comprehensive review of

optimisation methods targeted at practising naval architects. Rather than focussing
on solutions to specific problems, this work is primarily a description of
optimisation techniques, with special focus on the Nelder-Mead Simplex method
(Nelder and Mead 1965) and the Hooke-Jeeves direct search method (Hooke and
Jeeves 1961).
1.4.2 Analysis methods
Early researchers such as Nowacki et al. (1970), Jagoda (1973) and Kupras (1976),
based their optimisation work on statistical models derived from either the
performance of full-size ships, or measurements determined from tank testing. In
the case of resistance, regression analysis methods such as Holtrop and Mennen
·9·
(1978), Oortmerssen (1971) and Savitsky (1964) have been widely used in ship
design since their initial formulation, and subsequently in early optimisation studies.
The seakeeping optimisation examined in Bales (1980) used a regression model that
was based on measurements derived from a systematic series of hulls and analysed
using strip-theory calculations.
As low-cost computers became available in the early 1980s, direct numerical analysis
methods became more widely used in ship design optimisation. Ship resistance was
typically calculated using relatively simple numerical methods such as thin ship
theory (Hsiung 1981; Hsiung and Shenyan 1984; Hsiung and Shenyan 1985), with
seakeeping being predicted using strip-theory analysis of geometry based on Lewis
mapping of hull sections, (Walden et al. 1985).
Slender ship theory (Noblesse 1983) has also been used by several researchers. Due
to its performance advantages, slender ship theory continues to be of value in
circumstances where slender hulls such as catamarans and trimarans are optimised,
(Yang et al. 2000; Hendrix et al. 2001; Percival et al. 2001; Yang et al. 2001).
Computational Fluid Dynamics - panel methods
CFD is the process of simulating the fluid flow around a body such as a yacht hull
by solving a set of equations, ideally the Navier-Stokes equations, which describe
how the velocity, pressure, temperature and density of a moving fluid are related.
Due to the difficulty of finding solutions to these equations, CFD methods initially
used idealisations that simplified the task. These idealisations are -
• removal of viscosity terms from the Navier-Stokes equations yields the Euler
equations;
• removal of vorticity from the Euler equations yields the full potential
equations;
• linearisation of these yields the linearised potential equations.
These linearised potential equations were the first to be solved, with two-
dimensional methods being developed in the 1930s. However, a practical three-
dimensional method to solve the linearised potential equations was not developed
until 1966, (Hess and Smith 1966). This method breaks the surface geometry into
panels, resulting in this class of programs being referred to as panel methods.
Although panel methods have been used for some time for ship resistance
calculation, the first appearance in the literature of an optimisation process based
on panel methods is Maisonneuve (1993). Maisonneuve used the REVA CFD code
(Maisonneuve 1989) for the calculation of resistance, together with the AQUA code
(Delhommeau 1987) for seakeeping, in order to optimise a Small Waterplane Twin
Hull (SWATH) vessel.
· 10 ·
Advanced CFD methods in optimisation
Some optimisation work has taken place using Reynolds Averaged Navier-Stokes
(RANS) CFD methods. However, the high computational cost of RANS methods
has resulted in several different approaches being taken to reduce this burden.
An example of the performance penalties incurred by modern RANS codes is

provided by Parolini and Quarteroni (2007), who described analysis work
performed on ACC appendages for the Alinghi team for the 2007 America’s Cup.
Running on a cluster of 32 AMD Opteron processors, each simulation required
approximately 30 hours to reach convergence for a grid of up to 20 million points.
These very large computational grids were necessary due to the requirements
imposed by the transition and turbulence models that were used. Similarly, Korpus
(2007) described a RANS based hydrodynamic analysis of ACC yachts that took 8
days to analyse 147 simulation cases using 16 Itanium-2 processors for a grid of 6-8
million points.
Clearly, a global optimisation method requiring hundreds or thousands of RANS

iterations would be impractical. Although accelerating the performance by using
computing hardware with hundreds or thousands of processors is theoretically
possible, the cost for such a system is currently prohibitive.
One innovative approach to efficient optimisation using high quality solvers was
taken by Janson and Larsson (1996) with the SHIPFLOW code. This program
divides a ship hull into three zones, shown in Figure 4, applying a different
computational approach to each zone:
• Zone 1 uses a free-surface potential flow method for the whole hull and part of
the free surface.
• Zone 2 encompasses the thin layer close to the hull where a boundary layer
method of the momentum integral type is used.
• Zone 3 covers the aft portion of the hull and extends about one half of the
length of the ship downstream, as well as extending radially for about the same
distance. This zone uses a time averaged Navier Stokes solution to calculate the
effects of a thick, turbulent boundary layer where boundary layer methods
would fail.
Figure 4. SHIPFLOW zones, from Janson and Larsson (1996)
· 11 ·
The result is a program that retains sufficient accuracy for overall resistance
calculations, yet avoids spending unnecessary time on calculations that are not
required for particular portions of the hull.
Janson and Larsson used a gradient based method for their optimisation work, the
Method of Moving Asymptotes (MMA) that worked well for that particular
application. The method used to model and constrain the hull shape did not appear
to fare so well. The approach taken moved surface points along a vector, typically
the surface normal. However, no constraints were applied for flat of bottom, flat of
side, or longitudinal and transverse fairness, all essential requirements in the design
of a merchant ship hull.
Peri et al. (2001) investigated the use of variable complexity modelling, where
models and solvers of different levels of complexity are used at different stages of the
optimisation in order to improve performance. For example, during the early stages
of the optimisation process a 2D strip theory potential flow solver is used with a
simple geometrical model. As the optimisation proceeds, this model is defined in
greater detail and analysed, firstly using a full 3D potential flow solver, followed by
a RANS code utilising a multigrid, multiblock, finite-volume solver.
Campana et al. (2006) described several innovations designed to improve the

performance of what he refers to as Simulation Based Design (SBD) and set up two
different optimisation approaches to compare the alternatives. Campana makes
special comment on the intrinsic parallelism of Genetic Algorithms and their
suitability for implementation on multiple processors. This is seen as a way of
significantly accelerating the performance of RANS codes used in conjunction with
global optimisation methods.
In Campana’s first optimisation approach (labelled SBD-A), a Genetic Algorithm

having a very narrow available range for its independent variables is used. This
narrow-band GA was run in parallel on a 64 processor cluster.
Campana’s second optimisation approach (SBD-B) consisted of a variable fidelity

solver that used a multigrid method to trade off result accuracy against execution
time. Solutions in the early part of the optimisation process used a low resolution
grid, with the grid resolution increased as the solution was judged by a heuristic
procedure to be near to the optimal solution.
Note that Campana’s variable fidelity method differs from Peri’s variable complexity
method; variable fidelity modelling uses the same solver but diffent grid resolutions,
whereas variable complexity modelling also uses solvers having different levels of
complexity in the fluid-flow equations they solve.
Validation was performed, not only comparing original and optimised designs using
RANS analysis, but by also constructing tank models and measuring their resistance
· 12 ·
in a towing tank. Camapana obtained excellent correlation between the gains shown
by the RANS optimised design over its parent model and a corresponding
comparison of tank test model results. Seakeeping performance was also specified as
an inequality constraint, with the result that both heave and pitch RAOs showed
small improvements.
The two methods adopted by Campana achieved similar results in similar periods.
Both methods took approximately four days to run. However, SBD-A used 64
processors, while SBD-B used a single processor. These results suggest that there is
potential for combining the use of parallel genetic algorithms with variable fidelity
modelling to improve performance.
Metamodels
In cases where direct numerical analysis is prohibitively slow, a statistical

approximation, or metamodel, may be created in order to provide rapid evaluation
of an otherwise expensive objective function. Metamodels typically differ from older
regression models such as those proposed by Holtrop (1978) Oortmerssen (1971)
and Savitsky (1964) in several regards.
Metamodels are usually based on deterministic analytical methods and, as a result,

do not have the high levels of random noise often encountered with tank test data.
Although metamodels may use traditional polynomial regression, they are more
likely to use methods more suited to automated data analysis and fitting, such as
Response Surface Methods (RSM), (Minami and Hinatsu 2002), Kriging, Neural
Networks (NN) or Radial Basis Functions (RBF), (Peri and Campana 2005b; Peri
and Campana 2005d). These methods are better able to handle non-linear datasets
of high dimensionality and offer the possibility of creating a global approximation
model that is valid for the whole design space.
Neural networks and radial basis function networks have attracted a great deal of
interest for metamodelling. Within the domain of naval architecture, neural
networks have been used to create metamodels for ship resistance (Duvigneau and
Visonneau 2002), propeller design (Neocleous and Schizas 1995; Mesbahi and Atlar
2000), wetted surface area (Koushan 2003), hull form design (Mesbahi and Atlar
2000; Islam et al. 2001), catamaran resistance ((Couser et al. 2004; Mason et al.
2005)), ship stability (Alkan et al. 2004), roll stabilisation (Birmingham et al. 2002),
manoeuvring (Hess et al. 2004; Seif and Jahanbakhsh 2004), ship motions
identification and engine control (Mesbahi and Atlar 2000).
Several researchers have used metamodels to improve the performance of

optimisation or analysis methods that suffer from long calculation times. Peri and
Campana (2005c) used a variable fidelity approach in which a lower fidelity
· 13 ·
metamodel was used in the early stages of the optimisation, derived from a small
number of points calculated by a RANS solver.
As the optimisation progressed, the number of points used to calculate the

metamodel was increased, particularly in the regions of interest highlighted by the
optimiser. In the final stages of the optimisation a switch was made to a local
optimisation method that bypassed the metamodel and accessed the RANS code
directly.
This hybrid variable-fidelity approach, combining the use of a metamodel in the

early stage, with direct analysis in the final stage, provided an environment where
both fast global search and accurate local refinement could take place in an efficient
manner and shows great promise for future development.
1.4.3 Optimisation methods
Prior to 1997, the most widely used optimisation methods for naval architecture
were Sequential Unconstrained Minimisation Technique (SUMT), Sequential
Quadratic Programming (SQP), Nelder-Mead Simplex and Hooke-Jeeves direct
search, (Nowacki et al. 1970; Parsons 1975; Kupras 1976; Hsiung and Shenyan
1985; Walden et al. 1985; Keane et al. 1991).
SUMT and SQP are classified as gradient based methods, as they use derivatives of
the solution surface to determine the direction of each step in the optimisation
process. The Nelder-Mead and Hooke-Jeeves methods are considered pattern search
or derivative free methods, as they use direct comparisons between multiple points
to determine the direction taken for each successive step in the optimisation process.
Each of these methods can be considered a local optimisation method in that they
may become trapped in false optima if the solution space is multi-modal. The
widespread adoption of these methods by ship-design optimisation researchers
implies that they considered their optimisation problems to be unimodal; that is,
having only one potential optimum. However, later experience with global
optimisation methods has shown that this is not a valid assumption.
During this early period, multiple objectives were typically handled using weighted
sums of the individual objectives, leaving the outcome of the optimisation
dependent on the weights chosen. Alternatively, a measure of merit could be used
that was based on the sum of the squares of the differences between the preferred
values for each objective and the value obtained for the current design in the
optimisation process, as in Sarioz et al. (1992).
· 14 ·
Stochastic optimisation methods
During the early 1990s, significant experimentation with different optimisation

methods took place in the field of aerodynamics. One research direction was the
investigation of stochastic, global optimisation methods, such as Genetic
Algorithms (GA), (Goldberg 1989) and Simulated Annealing (SA), (Kirkpatrick et
al. 1983), as an alternative to the gradient based or pattern search based local
optimisation methods that had previously been favoured. The success of these
methods when applied to complex aerodynamic problems, (Quagliarella and
DellaCioppa 1994; Yamamoto and Inoue 1995a; Galan et al. 1996), suggested that
they might be equally applicable to optimisation problems in hydrodynamics.
Investigations into their use in naval architecture soon followed.
Sahoo (1997) compared two optimisation methods, Sequential Unconstrained

Minimisation Technique (SUMT) and Simulated Annealing (SA), using objective
functions derived from two statistical regression methods for planing hulls, Savitsky
(1964) and Keuning et al. (1993). Sahoo found that the SUMT method did not give
good results, being sensitive to initial parameter values, whereas the SA method was
found to be slow but robust. This was an early indication that stochastic methods
might be superior to deterministic methods when applied to non-linear functions
such as ship resistance.
The ability of stochastic global optimisation methods to solve problems with non-
linear and multi-modal solution spaces makes them particularly appropriate to the
optimisation of multihull designs. The interaction of the wave patterns between
hulls contributes in a non-linear manner to the resistance of multihulls, making
their resistance more difficult to predict.
Recognising this, Hearn and Wright (1997) used a GA to optimise a catamaran,

apparently the first instance of a GA being used for the optimisation of marine hull
shapes. Hull shape variation was achieved using a Lackenby transformation
(Lackenby 1950) of an existing hull shape. Fifteen different objective functions
were examined being various combinations of frictional resistance, wave making
resistance, added resistance, relative bow motion and vertical bow acceleration.
These values were calculated using the Salvesen, Tuck and Faltinsen (1970) strip
theory method for sea-keeping analysis, a Michell (1898) thin-ship theory based
wave-making resistance algorithm and a fine-form added resistance formulation.
The research performed by Day and Doctors (1997; 2000) shared many
characteristics with the work of Hearn and Wright. The target of the optimisation
was a catamaran, with both resistance and seakeeping measures included in the
objective function. A GA was also used to perform the optimisation. Hull geometry
was handled by a wireframe representation that modelled below-water hull shape
only, controlled by fifteen geometric variables.
· 15 ·
Both of these research projects resulted in workable optimisation systems using

Genetic Algorithms, although Day and Doctors were concerned at the limited range
of applicability of the analysis due to their reliance on thin-ship theory. Both
projects were limited by the use of simplistic methods for representing and varying
hull geometry, as well as by the use of simple weighted sums of multiple objectives.
Yasukawa (2000) compared a GA to the Hooke-Jeeves method for merchant ship

hull shapes, finding that the GA gave superior results. Similarly, Hirata (2004)
compared optimisations of merchant ships performed with a GA to optimisations
performed with SQP. Results showed that the SQP method consistently converged
to local optima rather than the global optimum, while the GA reliably found a
design that was close to optimal.
Hirata’s results are illustrated in Figure 5, showing the large number of points
evaluated by the GA, together with the sub-optimal path followed by the SQP
method.
Figure 5. Superiority of GA versus SQP, from Hirata (2004)
In this case the variables y1 and y2 relate to transverse waterline beam at particular
longitudinal locations, and are measures of the shape of the ship’s waterplane and
LCF. It can be seen that the SQP method converges on a local optimum that has a
higher coefficient of resistance relative to the best design found by the GA.
In recent years, other global optimisation methods have been investigated (Pinto et
al. 2004; Pinto and Campana 2005; Pinto et al. 2007) with positive results. These
methods included the Multistart Gradient Method, the Diagonal Rectangular
Algorithm for Global Optimisation (DRAGO), Particle Swarm Optimisation (PSO)
and Multi-Objective Deterministic Particle Swarm Optimization algorithm
(MODPSO).
The Particle Swarm Algorithms (PSO and MODPSO) performed well, particularly
for problems involving multiple objectives, with significantly improved results
compared to the local optimisation methods examined (Nelder-Mead Simplex and
Hooke-Jeeves pattern search), although at the cost of increased calculation times.
· 16 ·
Multi-objective optimisation
Although there have been many efforts to optimise multiple objectives using
weighted sums of output values, true multi-objective optimisation of marine vessels
was first described by Sen and Todd (1997) and Poloni and Pedirodav (1997). In
both cases, the optimisation procedure was a Multi-Objective Genetic Algorithm, or
MOGA, provided by the modeFRONTIER optimisation shell. Rather than the
output from the optimisation being a single optimal design, the MOGA located a
collection of designs that occupied a Pareto optimal front.
The Pareto front is made up of those designs that are better than all other designs in
at least one objective measure; that is, they are non-dominated designs. For example,
in a problem where larger values of functions and are better, as illustrated in
Figure 6, from (Zitzler 2002), the non-dominated points on the Pareto front are
those that are better than all other points in either or .
Figure 6. Illustration of Pareto optimality, from Zitzler (2002)
Multi-objective methods are necessary in situations where there it is not possible to

frame the optimisation problem as having a single objective with multiple
constraints. However, if it is possible to reduce the problem to having a single
objective, this is to be preferred, as using multiple objectives means that a subjective
choice remains for the designer at the conclusion of the optimisation process.
Multidisciplinary optimisation
Multidisciplinary optimisation differs from multi-objective optimisation in that the

key objectives do not share common features. For example, the optimisation of the
effects of hull shape variations on resistance and seakeeping uses multiple objectives
that are hydrodynamic in nature. These can potentially be calculated using a single
piece of software. On the other hand, the simultaneous optimisation of the
resistance, structural strength and cost of a vessel involves calculations drawn from
three unrelated disciplines.
Neu et al. (2000a) proposed a multidisciplinary approach to the optimisation of

commercial ships. The objective function for this optimisation was a measure of
required freight rate, determined by combining conventional measures of
performance with cargo capacity and ship economics.
· 17 ·
Poloni et al.(1999) used a MOGA to perform a multidisciplinary optimisation the

keel of a sailing yacht. This involved optimising lift and drag while constraining keel
depth, mass, centre of gravity and structural strength. These factors present a
complex trade off between a thinner foil section with smaller volume having less
resistance, versus a thicker foil having greater strength and righting moment.
This particular problem is directly applicable to ACC yacht design as ACC keels
have significant problems with lateral deflection due to their small section coupled
with the 18 tonne lead bulb attached to the tip of the keel fin. The trade off
between performance loss due to deflection of the keel under load, versus the
increase in hydrodynamic drag due to increased keel thickness, is an optimisation
problem that was addressed for ACC yachts by Campana et al. (2007).
Robust design optimisation
Whitfield (1998; 1999) focussed on the optimisation of catamaran configurations

using a genetic algorithm. In this case, the measure of merit was solely concerned
with ship motions. However, the major contribution of this work is the
introduction of the concept of robustness to the field of marine design. Whitfield
defines robustness as –
“A product’s robustness is a measure of the variation in its utility experienced in a

typical application. That is to say, the lower the sensitivity or variation in utility,
the greater the robustness of the design.” (Whitfield et al. 1998, p.373)
The problem of robustness is often overlooked, with the result that the optimised
designs work well only in a given environment and perform poorly when the
conditions fluctuate slightly. Such a design is likely to be less desirable than one
whose peak performance is lower, yet performs well in a variety of conditions.
This situation can be observed in many of the attempts to optimise ship hull lines
for a single target Froude number (Fr), such as Hsiung (1981), Janson and Larrson
(1996), Dejhalla et al. (2001) and Chen (2004). If fairness and convexity are not
constrained, the result is usually a hull shape with longitudinal undulations, as
shown in Figure 7.
Figure 7. Longitudinal undulations in hulls optimised for a single

speed, from Chen (2004)
· 18 ·
Such a shape may be ideal for one particular speed, but performs poorly if the speed
is changed only slightly. Percival et al. (2001) optimised the hull shape of a Wigley
hull at three different Froude numbers, resulting in three very different hull shapes,
as shown in Figure 8.
Each of these hulls had significant concavities in the longitudinal direction, as well
as distinct bulbs at the bow and stern, which were quite extreme for the hull
optimised for Fr = 0.408. However, when the parent hull was optimised for the
three Froude numbers simultaneously, the resulting hull shape had long fine ends
with no trace of bulbous bow or stern, or longitudinal undulations.
Figure 8. Optimisation of a Wigley hull at different Froude numbers,

from Percival et al. (2001)
Figure 9 shows curves of total drag coefficient CT for the parent Wigley hull, the
three hulls optimised for single Fr values, and the hull optimised for all three Fr
values. The graph shows how poorly the three hulls optimised for a single Fr
perform in off-target conditions, with resistance increasing rapidly as Fr deviates
from the target speed. Conversely, the hull optimised for a combination of three Fr
values has resistance values only slightly greater than the three single-speed
optimised hulls at each of their target speeds.
Figure 9. Resistance curves for hulls optimised by Percival et al. (2001)
· 19 ·
The concept of robustness is directly applicable to sailing yacht design. A boat that
performs exceptionally on one point of sailing, in one set of weather conditions, is
generally not as desirable as a boat that performs well in a variety of wind velocities,
sea conditions and points of sail. A measure of merit formulated for the
optimisation of a racing yacht needs to take into account a wide variety of
conditions, rather than focussing on a narrow range of heel, yaw and velocity values.
1.4.4 Hull shape variation methods
In order for a design optimisation process to function, a method for parametrically

varying individual designs is required. This is not a trivial problem, as not only is it
necessary to vary hull shape based on changes to parameters such as CP and LCB, it is
also necessary to ensure that constraints, such as constant displacement and hull
fairness, are not violated.
Parametric transformation
Initial attempts at hull design optimisation such as Hearn and Wright (1997) used
the method described by Lackenby (1950) to vary hull shape. Lackenby’s method
performs a longitudinal shift of hull sections in order to modify parallel mid-body,
CP and LCB.
Although Lackenby’s method is good at retaining the fairness and features of the
parent hull, the basic method allows displacement to vary in an uncontrolled
manner. This must be compensated for by scaling the hull shape along its principal
directions. A more significant problem is that Lackenby’s method does not provide
any mechanism for varying the cross sectional shape of the vessel.
Markov and Suzuki (2001) applied an analogous approach to a the control vertices
of a B-spline surface representing commercial ship hull forms. Hull shapes were
optimised using the Davidson-Fletcher-Powell (DFP) method with resistance
calculated using a higher-order Rankine source panel method.
Although other methods have been proposed to permit more flexible variation of a
parent hull to match desired longitudinal and cross sectional parameters, such as
those by Söding and Rabien (1977) and McNaull (1980), these approaches do not
appear to have been utilised in any optimisation research.
· 20 ·
Hull blending
One alternative shape variation method was suggested by Chen and Parent (1989),
who explored the possibility of using a weighted average of two or more parent
models. Although this research was aimed at general industrial design, the approach
is equally applicable to naval architecture.
A similar approach was implemented by Neu (2000a; 2000b; 2007), who used a
barycentric blend of a number of different hull shapes within an optimisation
procedure. This can be represented as:
Resultant Ship Hull Basis Hull
where:
Blending Coefficient for Basis Hull
and
0 1,
1, 2, … , N
Figure 10 illustrates a simple midship section formed using this method by blending
two basis hulls.
Figure 10. Barycentric blending of hull shapes, from Neu et al. (2000)
This approach is particularly easy to implement if the parent hull representations are
NURBS surface models, made up of control point networks having the same
number of control points. In this case, a weighted sum of each of the control point
co-ordinates from each parent hull surface is used. This method has been used by
ACC designers since 1990, (D. Peterson 1992, pers. comm.) for the creation of
systematic series of hull models.
The disadvantage of the blending technique is that individual design parameters are
not independent, as all parameters are fully correlated with the blending coefficient.
This results in large portions of the design space being inaccessible to the
optimisation procedure.
· 21 ·
Parametric modelling
Parametric modelling, the generation of a ship hull from scratch based on a set of
key parameter values, was pioneered by Nowacki (1970; 1977; 1983; 1993; 2005)
with contributions from Walden et al. (1985), Peacock (1996), Birmingham and
Smith (1998), and Islam (2001).
However, it was not until the work of Stefan Harries that parametric modelling of
ship hulls became a viable alternative to the parametric transformation of parent
models for marine hull design and optimisation.
Harries, in collaboration with Claus Abt and others, has been possibly the most
prolific investigator on the subject of hydrodynamic optimisation of marine vessels,
(Harries and Abt 1999a; Birk and Harries 2000; Abt et al. 2001a; Harries et al.
2001b; Valdenazzi et al. 2002; Abt et al. 2003; Harries et al. 2003b; Heimann and
Harries 2003; Valdenazzi et al. 2003; Abt et al. 2004).
The common component in these works is the use of the FRIENDSHIP-Modeller,

later known as the FRIENDSHIP-Framework, to parametrically define and vary hull
geometry. Developed by Harries and Abt, the FRIENDSHIP tools provide a flexible
environment for hull shape creation and manipulation, while retaining high quality,
fair hull surfaces.
The work of Harries and Abt has covered a wide variety of vessel types, including
semi-submersible offshore rigs, military vessels, fast ferries, and containerships.
Optimisation methods used have included Hooke-Jeeves direct search, the tangent
search method, Newton-Raphson solvers, Multi-Objective Genetic Algorithms
(MOGA) and the Broyden-Fletcher-Goldfarb-Shanno (BFGS) method.
This extensive body of work demonstrates the utility of the parametric modelling
approach for a wide range of hull types and analysis methods. However, the fact that
this method does not use a known parent model as a starting point, nor is it able to
exactly reproduce such a parent model if required, is a potential stumbling block in
the field of ACC yacht design. ACC designers wish to have specific features and
preferences incorporated into a hull shape and a system that is not able to cater to
these requirements is at a significant disadvantage.
free-form deformations
Sederberg and Parry (1986) proposed a method for deforming geometric objects by
embedding them within a flexible volume defined by trivariate splines. This
enclosing volume could then be deformed by moving its defining control points,
resulting in an equivalent deformation to the embedded object.
· 22 ·
FFDs have been used successfully for aerodynamic optimisation (Desideri et al.
2006; Duvigneau 2007) as they permit smooth deformation of complex geometries
using a small number of variables.
Figure 11. An FFD used for deformation of an aircraft wing (© INRIA

Opale Project-Team 2007, reproduced with permission)
Peri and Campana (2005a) examined the use of the Free Form Deformation (FFD)
in ship design optimisation. In this case, FFDs were used to enclose and deform
specific blocks of the ship, namely the bulbous bow, forward waterlines and stern
section, in an effort to reduce the resonant whipping behaviour of the ship. The
FFD used to deform the bulbous bow of the ship is shown in Figure 12, with
control points subject to movement shown with black arrows.
Figure 12. An FFD lattice enclosing the bulbous bow of a cruise ship,
from Peri and Campana (2005a)
Duvigneau and Visonneau (2001; 2002) investigated the use of FFDs for the
deformation of ship hull forms for optimisation. Duvigneau showed that FFDs
provided a flexible way of varying design geometry and resulted in significant time
savings when the FFD was used to directly deform the computational grid used for
CFD analysis, rather than recreating the grid at each iteration.
The optimisation systems incorporating FFDs that were used by these researchers
were implemented successfully. However, the FFD method is not trivial to
implement and there appears to have been little further work utilising FFDs within
the field of marine design optimisation.
· 23 ·
FFDs and evolution
One inspiration for the use of the FFD method in design optimisation is drawn
from parallels with biological evolution. D’arcy Thompson’s “On Growth and
Form”, (Thompson 1917), detailed the variations that occurred during the
evolution of different species. Thompson showed that many of the high-level
changes that occurred in evolution consist of both linear and non-linear
transformations of previous forms (Figure 13).
Figure 13. Non-linear deformation of primate skulls, from Thompson (1917)
Thompson’s work also illustrated some of the different types of deformations that
may take place. Figure 14 illustrates various linear, radial and non-linear
deformations that Thompson suggested could describe some of the variation that
occurs between different species of fish.
Figure 14. Variation of fish forms using linear and non-linear transformations, from Thompson (1917)
Each of these mappings is directly analogous to transformations required to deform

a parent hull shape to match a set of desired hull parameters.
Both natural evolution and evolutionary optimisation systems benefit from methods
that can achieve significant and effective variation, yet can be encoded in a concise
manner. Biological evolution is able to vary small details as well as apply global
variations, encoding each efficiently using a small number of parameters. Such a
hierarchy of transformations may also be of value when applied to the variation of
geometry during an optimisation task, reducing the number of dimensions in the
search space and allowing the optimisation to proceed more efficiently.
· 24 ·
1.4.5 Sailing-yacht design optimisation
Although there has been a great deal of research performed on the calculation of
yacht performance, there is a relatively small body of literature specifically on the
subject of computer based yacht design optimisation.
An early use of genetic algorithms in sailing yacht design was described by Day
(1993). In this work, the sail plan of a yacht was optimised using an objective
function based on a non-linear vortex lattice model. Once the GA had converged on
a solution, a further optimisation was performed using a search pattern based
algorithm, Powell’s direction set method. In all cases, the classical optimisation
method was unable to improve on the solution found by the GA. This result shows
that genetic algorithms are capable of converging on solutions of equivalent quality
to traditional optimisation methods when faced with real-world problems. However,
Day commented that performance of the GA was a significant issue.
Poloni et al. (1999) used a Multi-Objective Genetic Algorithm (MOGA) combined

with a neural network metamodel to optimise the lift and drag of the keel of a
sailing yacht while constraining mass, centre of gravity and structural parameters.
Limitations of this work were the use of a fixed heel and yaw angle, rather than a
simulation of a variety of sailing conditions and keel loadings.
Other than the yacht design optimisation work performed by Harries, Abt and
Hochkirch (Harries and Abt 1999b; Abt et al. 2001b; Harries et al. 2001a;
Hochkirch et al. 2002) there has been little other yacht design optimisation work
performed outside of the America’s Cup arena.
Optimisation work that has been performed specifically for the design and
refinement of Americas Cup yacht design (Harries et al. 2001a; Philpott 2003;
Philpott et al. 2004; Baik and Gonella 2005; Peri and Mandolesi 2005) is discussed
in detail in Chapter 2.
· 25 ·
1.5 Design optimisation summary

A review of the literature concerning the automated optimisation of marine vessels
has revealed several clear trends over the past thirty years, illustrated in Table 2.
Table 2. Summary of Developments in Yacht Design Optimisation
Hull Variation Hull Measure of Optimisation Analysis Method

Method Geometry Merit Method
Variation of Parameter Single objective Gradient based, Polynomial

parameter values values only local optimisation regression of tank
test data
Lackenby Wireframe Multiple Derivative free Thin ship/ slender

← Time
objectives pattern searches ship theory
Parent Blending, Polyhedral Multi- Simulated Panel methods,

Parametric disciplinary Annealing potential flow
Modelling optimisation codes
Free Form NURBS Robust design Stochastic global RANS codes

Deformation optimisation optimisation
Each of these developments has involved greater algorithmic complexity and a larger
number of calculations compared to its predecessors. This has been offset to some
extent by hardware developments, both the inexorable rise in CPU performance in
accordance with Moore’s law, (Moore 1965), which predicts a doubling of CPU
transistor count every 24 months, as well as the increasing use of multi-core chips
and parallel clusters for computationally intensive applications.
Despite the exponential increase in computing power since the introduction of the
transistor, this has not been as great as the rate of increase in the number of
calculations demanded by modern analysis and optimisation methods. This
imbalance between processing speeds and processing needs has resulted in the
exploration of approaches that can streamline the optimisation process and reduce
the total computational cost. One relatively recent addition to the field of
optimisation is the adoption of a metamodel to act as an inexpensive surrogate for a
computationally expensive objective function. Although not applicable to all
situations, pre-calculated metamodels may dramatically improve the performance of
some ship-design optimisation problems.
The results of these developments have been encouraging, with a genuine

improvement in resistance and seakeeping performance for many vessels.
Importantly, automated optimisation methods have also been responsible for the
discovery of novel and non-intuitive hull shapes, such as those reported by Harries
et al. (2006). Harries demonstrated that automated optimisation methods may
occasionally be able to uncover superior combinations of design parameters that
might have otherwise been overlooked or dismissed by a human designer.
· 26 ·
1.6 Outline of thesis

One of the key motivations for the work described in this thesis is the conflict
between the need for accurate calculations and the desire to get results within an
acceptably short time. In order for the results of an optimisation to have anything
but academic interest, it must be able to make design recommendations that result
in real improvements when verified with independent CFD analysis, tank testing
and full sized trials. Yet these results must also be produced within the restricted
budget and limited period of time available during the design cycle of an America’s
Cup campaign.
In order to create a workable optimisation system, several key problems must be

solved. These include -
• The selection of a measure of merit that is appropriate to the problem of
determining the best yacht to compete in a match racing series, against an
opponent whose design parameters are unknown. This question is addressed in
Chapter 2, Measure of Merit.
• The creation of parametric transformation methods that allow ACC hulls to be
varied to match specific parameters, yet satisfy the constraints imposed by the
ACC rule (ACC 2003). These areas are covered in Chapter 3, Constraints, and
Chapter 4, Data Approximation.
• The adoption of techniques to enable an otherwise expensive set of analysis
functions, such as CFD analysis of the flow around a yacht hull, to be rapidly
evaluated. This is necessary in order for the optimisation environment to have
acceptable performance. This subject is discussed in Chapter 4, Data
Approximation.
• The selection of CFD method for determining the lift and drag of a hull shape
at various heel angles, yaw angles and velocities is discussed in Chapter 4, Data
Approximation. The selection or creation of Velocity Performance Prediction
(VPP) and Race Modelling Program (RMP) software for determining and
comparing the sailing performance of different designs is covered in Chapter 5,
Performance Evaluation.
• The selection of an optimisation method capable of locating an optimal design,
in what is likely to be a non-linear, multi-modal solution space. The method
should be suitable for use with the chosen measure of merit. The choice of
optimisation method is discussed in Chapter 6, Optimisation Algorithm.
• Integration of all components into a robust system that produces reliable and
repeatable results, see Chapter 7, Implementation, and Chapter 8, Verification
and Validation.
• Finally, results obtained and their implications for future work are detailed in
Chapter 9, Results and Discussion, and Chapter 10, Conclusions.
· 27 ·
1.7 Contribution
Although the research described by this thesis focuses on a narrow range of racing-
yacht design parameters, the approach described is broadly applicable to a wide
range of marine craft. It is envisaged that this research will contribute not just to the
design of ACC yachts, but to the field of naval architecture in general.
A significant amount of recent research exists involving the optimisation of ships

and boats. However, much of this is focussed on the use of direct CFD solvers
having prohibitively long run times. The approach taken with this research,
combining the use of a high-fidelity meta-model with a stochastic global
optimisation method utilising a probabilistic measure of merit, has the potential to
allow optimisations, which would otherwise be impractically slow, to be completed
within a reasonable time on conventional personal computers.
This research has identified problems and created innovative solutions in the several
areas. Contributions to the field include:
• A survey of the state of the art for the design optimisation of marine vessels.
• The design and implementation of a geometrical parametric transformation
function based on a hierarchical free-form deformation procedure.
• The use of neural network metamodels for hydrodynamic data for ACC yachts;
• The identification of an appropriate measure of merit for an ACC design
optimisation process, and the creation of an appropriate race and tournament
modelling system to support this.
• The modelling of the random variation that occurs in the yacht-racing
environment.
• Recognition of the dynamic fitness landscape that results from the adoption of
a competitive fitness function, and the identification of its effects on the
choice of optimisation method.
• Recognition of the potential for a co-evolutionary arms race to occur, due to
the use of the competitive fitness function, and the implementation of steps to
limit any adverse effects from such an arms race.
• The treatment of the ACC design optimisation problem as a stochastic,
multiplayer game, and recognition that this system may have no Nash
equilibrium, due to penalties incurred at each turning mark.
· 28 ·
Chapter 2 · Measure of Merit
2  Measure of Merit
The first step in setting up an optimisation process is to precisely determine the
objective of the optimisation. This is often not a trivial exercise, with multiple
candidates for the measure of merit presenting themselves. In many cases, what
initially appear to be objectives can be treated as constraints; in other cases, multiple
objectives remain, and these require the adoption of multi-objective optimisation
methods.
Many attempts at yacht design optimisation have viewed the task as a multi-
objective problem, or at least a single objective problem with a large number of
constraints. Although this may be a reasonable assumption for cruising yacht design,
where there are conflicting requirements for speed, comfort, stability, seaworthiness
and cost, ACC optimisation is based solely around the wish to be the winner of a
specific set of races, held at a particular location during a specific month and year.
As a result, the design optimisation of ACC yachts can be regarded as a problem
with a single objective, although the formulation of the objective function is a
formidable task.
· 29 ·
2.1 Measure of merit or objective function?

Although the terms “measure of merit” and “objective function” are generally
regarded as synonymous in optimisation research, this thesis distinguishes between
the two terms.
The reason for making a distinction is simple. In most optimisation procedures, the
measure of merit used to discriminate between designs is a deterministically
calculated value; that is, given the same set of input parameters, the output value
will be the same each time it is calculated and as such is truly an objective measure.
For some optimisation procedures, this is not the case. The worth of a particular
design may be measured, not against some objective criteria, but relative to the
performance of another design. Depending on the performance of the design being
used for comparison, the current design point in the optimisation may be ranked
comparatively high or low. Consequently, the measure of merit for a specific set of
design parameters may differ during the optimisation process.
Within this thesis, the term “objective function” is taken to mean an algorithm used
to calculate the performance characteristics for an individual vessel. These
calculations are purely deterministic and therefore can be considered an objective
measure of the performance of a yacht. The “measure of merit”, on the other hand,
is the final value, derived from the output of the objective function, used by the
optimisation procedure to rank designs.
There are several potential candidates for adoption as a measure of merit for an ACC
yacht design optimisation. The form taken by this measure of merit influences the
choices made at every stage of the optimisation process. It is therefore important to
closely examine the benefits and drawbacks of each possible alternative.
2.2 Upright resistance

Many researchers have attempted to minimise the resistance of an upright hull, with
or without appendages, at one or more speeds. While this approach has been used
primarily for the optimisation of ships, there have also been examples, such as
Harries et al. (2001a) and Maisonneuve (2003), where this approach has been
applied to ACC yachts.
In both cases, unappended hulls operating at zero heel and yaw for a single Froude
number were examined. However, for yachts that operate over a wide range of
velocities, heel and yaw angles, the upright case is not indicative of the overall
performance of the yacht, and this approach is of little value.
· 30 ·
2.3 Heeled lift and drag

A more detailed form of analysis encompasses the optimisation of both lift and drag
at multiple velocities, heel and yaw angles. This approach has mainly been used by
authors working directly with CFD solvers such as Peri and Mandolesi (2005).
Although this approach is an improvement over the upright case, it suffers from not
incorporating a measure of stability, and therefore sail carrying ability, in the
objective function. For example, a narrower hull may have less resistance, but may
also have less stability, giving it less sail carrying ability. There is no way of knowing,
without more-detailed analysis, whether the yacht would be faster or slower overall.
Baik and Gonella (2005) investigated the performance of ACC yachts using an
optimisation of the resistance of unappended canoe bodies at four different angles of
heel. Hulls were analysed with zero yaw and no account was taken of righting
moment. The results of these optimisation runs illustrate the pitfalls of using
simplistic measures of merit (Figure 15).
Figure 15. Optimal hull shapes for different heel angles, from Baik and Gonella (2005)
Each successive optimisation run used a design objective that differed by only 10° of
heel angle from the previous run, yet the optimised hulls are quite dissimilar, with
obvious wide variation in BWL, CM, TC, flare and beam at transom.
Despite their optimisation runs producing many different hull shapes, Baik and
Gonella did not provide any methodology for selecting which of the various designs
should be recommended for a particular America’s Cup match.
· 31 ·
2.4 Velocity Prediction Programs

For stability to be properly incorporated, it is necessary to combine lift and drag
data for the hull with information about the weight, centre of gravity and sail plan
of the yacht, in order to calculate actual sailing performance in different wind
velocities and different points of sail. Software that calculates yacht performance in
this way is referred to as a Velocity Prediction Program, or VPP. Although a VPP can
calculate a yacht’s speed, heel angle and leeway for a given wind velocity and
direction, overall performance is often presented graphically in the form of a set of
polar performance curves, as shown in Figure 16.
Figure 16. Polar performance curves, © ORC 2004, reproduced with permission
Much work has gone into the development of VPPs during the past thirty years,
(Kerwin 1978; Oliver and Claughton 1995), with the result that they are now
capable of producing reasonably accurate sets of polar performance data over a range
of true wind directions and wind velocities for a wide variety of yacht types.
Although the use of yacht polar performance data is superior to the use of hull lift
and drag as a measure of merit, it raises the difficulty of how one selects the exact
conditions for which the yacht is to be optimised. In some cases, such as Fassardi
and Hochkirch (2006), the best VMG for a single wind velocity is used as a measure
of merit. In other cases such as Jacquin et al. (2002), VMG for multiple wind
velocities are used for multi-objective optimisation.
· 32 ·
Jacquin et al. (2002) used upwind and downwind VMG values for 10 and 20 knots of
wind as objectives in a multi-objective optimisation of ACC yacht designs. This
analysis resulted in a group of designs being identified as Pareto optimal for the
specified conditions, as shown in Figure 17, but no guidance was given as to how a
selection should be made among these designs.
Figure 17. Pareto optimal ACC designs, from Jaquin et al. (2002)
Korpus (2007) described an optimising VPP that used RANS analysis of both
hydrodynamic analysis of the hull and appendages, as well as aerodynamic analysis
of the rig. Optimisation capabilities allowed up to four independent variables,
which could be design parameters such as beam, CP or mainsail camber, or
operating parameters such as traveller position or tab angle. Independent variables
could be restricted to the hydrodynamic or aerodynamic domains, or could be
mixed.
The measure of merit chosen by Korpus was VMG for multiple wind velocities.
However, how these were chosen and weighted relative to one another was not
described.
While these approaches are an improvement over simple lift and drag measures, VPP
output is not sufficient to differentiate between two boats unless the weather in
which the boats are to be sailed is taken into account. As stated by Oliver et al.
(1987):
“There is an essential stochastic character in yacht racing, in that the relative

performance of two yachts depends on the wind speed and sea conditions, which
vary randomly from day to day, and more predictably from month to month. VPP
results by themselves are therefore inconclusive and possibly misleading for
determining the order of merit of two candidate yachts” (Oliver et al. 1987,
p.240)
· 33 ·
2.5 Comparison plots

In order to compare the performance of two boats it is necessary to convert the VPP
derived polar data into comparison plots as shown in Figure 18.
Figure 18. Performance comparison plot
Wind speed is plotted on the x-axis and deltas between the VMG of the two boats on
the y-axis, typically expressed in metres per minute or seconds per mile. The
performance of the reference boat is plotted as a horizontal line and the test boat
deltas plotted against it, with points below the abscissa being faster and points above,
slower.
Comparison plots or time deltas are not appropriate as a measure of merit for an
automated optimisation process, as the plots contain no information about the
wind velocities or course directions in which the boats will be sailing. Hence, unless
one boat is superior in all wind velocities and at all apparent wind angles, the
possibility will exist that the outcome of the race will be dependent on the weather
and course.
In spite of this limitation, comparison plots have been widely used as the basis for
manual design optimisation for some time, (Chance 1987; Rosen et al. 2000;
DeBord et al. 2002). Most importantly, they provide a rapid visual check of overall
performance, and therefore play an important role in the validation of any designs
created by an optimisation system.
· 34 ·
2.6 Race Modelling Programs

A Race Modelling Program (RMP) uses performance characteristics derived from a
VPP for two or more boats, in order to simulate a race between the boats. In its
simplest form, the RMP may use a uniform wind velocity and direction and ignore
interactions between boats. However, in order to model an entire yacht race more
accurately other factors need to be taken into account.
On an actual racecourse the wind varies in both speed and direction, boats influence
one another with backwind and wind shadow, there are multiple legs and boats are
forced to concede right of way on the course and at rounding marks. The result is a
bias in favour of the boat that is faster upwind and can lead around the first
windward mark. Statistics collated by the author for the America’s Cup Acts
between 2004 and 2006 show that for the top four boats, 80% of races were won by
the boat that rounded the first weather mark in front. This corresponds with a
figure of 80% quoted in several articles (Lloyd 1995; Clarey 2000) for America’s
Cup racing in general.
To handle these conditions and more accurately account for this bias it is necessary
to create a more detailed simulation. This can be achieved by dividing the race into
discrete periods and stochastically sampling wind conditions for each step from
distributions derived from historical data.
Many RMP have been developed to date with varying levels of complexity, from
simple probabilistic methods through to fully detailed simulations. Possibly the
most sophisticated to date is the ACROBAT program, described by Philpott (2003;
2004). ACROBAT is a fixed interval time stepped simulation and was intended to
be a highly accurate model of the racing performance of an ACC yacht. ACROBAT
incorporated detailed calculations for many aspects of yacht racing including:
• A stochastic wind model generated using a Markov chain.
• Independent wind fields for each of the two yachts in the race, correlated
according to their spatial separation.
• Modelling the dynamics of both tacking and mark rounding.
• Interactions between yachts, including backwind effect and wind shadows.
• Route optimisation, covering and collision avoidance penalties.
However, choosing the level of detail required to rank two boats accurately for the
purposes of optimisation is not straightforward. Simulations that are more detailed
have longer execution times but may not necessarily have a better ability to rank the
performances of two boats in a match race.
· 35 ·
2.7 Monte Carlo race model simulations

Regardless of how sophisticated it is, a simulation of a single race does not provide
sufficient information to permit full optimisation of a yacht design. A single race
cannot sufficiently capture all of the random variation in conditions for which
racing yachts need to be designed.
A more effective measure is based on a Monte Carlo simulation, where the RMP is
run repeatedly for hundreds or thousands of races. Although Monte Carlo race
simulations do not appear to have been used for an automated design optimisation
procedure, there has been some excellent work done in this area for the comparison
of specific yacht designs. This work includes that performed by PACT, the
Partnership for America’s Cup Technology, (Gretzky and Marshall 1993) and the
ACROBAT program, (Philpott 2003; Philpott et al. 2004).
However, the seminal work on the use of Monte Carlo simulations for the design of
America’s Cup yachts was performed by the Sail America team for the 1987 Cup
(Chance 1987; Letcher et al. 1987; Oliver et al. 1987).
Racing to select the challenger for the America’s Cup final started in late October
when the winds were moderate and continued through to the end of January when
the sea breezes were very strong. The different rounds of the challenger selection
series, the Louis Vuitton Cup, were also awarded progressively more points for
successive rounds of competition, making early losses less important in terms of
total points. Boats that performed poorly were eliminated after certain rounds in the
competition.
The development of the winning yacht in the 1987 America’s Cup, Stars & Stripes
87, was recounted in Letcher et al. (1987).
“All the technology described (CFD, VPP, RMP and Monte Carlo racing
simulations) flowed together into a strategy for the design of Stars & Stripes 87. It
owes its very existence to the velocity-prediction and race-model programs, because
without their clear dictates it would surely not have been built.
Our earlier 12-Metre designs, two yachts named Stars & Stripes but further
designated as ’85 and ’86 according to the year of construction, had proved to be
of unprecedented size, power, stability and speed in heavy winds. Stars &
Stripes ’85 had proved to be slightly faster under most conditions, and its crew had
developed tremendous confidence in the boat.
In late 1985 and early 1986, however, when the Sail America team was training
and testing in Hawaii, the Australian defender candidates and many of the
challengers were training in Perth and competing in the 12-Metre World
Championship races. Careful observation, including photogrammetric analysis,
· 36 ·
showed that all these boats were substantially smaller than we had decided would
be optimal for summer conditions. Results from the race-model program showed
that although Stars & Stripes ’85 had high probabilities of beating any known
competitor in a four-of-seven series in January or February, it had only a
marginal chance of surviving the round-robin eliminations among a fleet of 13
challengers that were mostly two or three feet shorter.
Faced with these predictions, Sail America had no choice but to build another
boat small enough to compete effectively in the round robins. A new design could
take advantage of the new hull shapes (developed with slender-ship theory and
confirmed in tank tests) and of progress in computer optimised keel design. The
length was chosen to be a little more than the rest of the fleet, as suggested by game
theory. After the elimination rounds the span of the winglets could be increased,
the boat re-ballasted for greater weight (a move that would produce a longer
waterline) and the keel made more bulbous to lower the centre of gravity. Thus
equipped for the stronger winds of summer, the boat would be hard to beat.”
(Letcher et al. 1987, p.40)
The predictions of the Sail America design team proved correct. Stars & Stripes 87
struggled to win races in the early rounds of the challenger selection series, despite
significant changes to sail area and ballast aimed at “re-moding” her for light winds.
As the summer progressed and the Fremantle sea-breezes strengthened, Stars &
Stripes 87 improved her standing relative to the other challengers, and this
culminated in the boat winning the challenger selection series. Stars & Stripes 87
proceeded to dominate the Australian defender in the America’s Cup series, winning
four races to nil.
The work performed by Sail America was ground breaking for several reasons. Not
only did it result in an America’s Cup winning design, it did so by adopting an
apparently high-risk strategy that would not have been obvious without the analysis
performed using Monte Carlo and game theory based methods. Other design teams
faced with the same weather predictions concluded that a boat with good all-round
performance was required. In contrast, Sail America’s early analysis determined that
losses in light winds during the early rounds of the challenger selection series were of
little importance, compared to wins in the later, heavy weather rounds, as long as
Stars & Stripes 87 could avoid elimination.
The surprise for Sail America was that other teams did not come to the same
conclusion and build boats tailored to stronger winds. The realisation by Sail
America that its best yacht for the expected conditions was not ideal, because of the
designs chosen by its competitors, was crucial. The game-theory based analysis that
followed, which resulted in a compromise length being chosen which was optimal
compared to the known dimensions of competing yachts, was instrumental in Sail
America winning the 1987 America’s Cup.
· 37 ·
2.8 Tournament modelling

The research performed by the Sail America design team illustrates several key
points:
• Variations in wind velocity, both intra-day and across the duration of the
competition need to be accounted for.
• The structure of the tournament (i.e. number of competitors, points allocated
per race and timing of competitor eliminations) can affect the outcome of the
simulation.
• Most significantly, the ideal boat for a competition may not necessarily be a
static optimum; rather, it may be dependent on the design of the opposing
boats.
These points are as valid today as they were in 1987. The America’s Cup is now a
best of nine race series, while the Louis Vuitton Cup, the selection series for the
America’s Cup challenger, has significantly more races over a period of 8 weeks.
The successful challenger for the America’s Cup has to first win the challenge series
and then compete against the defending yacht. The total period for this is close to 3
months, over which the standard deviation of wind velocity will be significantly
higher than the typical standard deviation for a typical sailing day. Due to seasonal
variations, the mean wind velocity may also vary over the 3 month period.
In addition, the America’s Cup is a tournament of individual matches between two

yachts, and success is based on points accumulated for race wins, not accrued race
times. Races won by one second carry as many points as races won by five minutes.
This has a significant effect on the measure of merit. For example, compare a yacht
that wins five races by five seconds each and loses one race by one minute, against a
yacht that loses five races by five seconds and wins one by one minute. The second
yacht may have a shorter accrued time for all the races taken together, so arguably
could be termed the faster yacht. However, the first yacht will get the majority of
the points, and in the case of the America’s Cup final, would win the event.
To account for these effects it is important to evaluate the ability of a yacht to win
races against a range of opponents, over a range of weather conditions, by
simulating an entire tournament of races.
· 38 ·
2.9 Weather model

The weather model for VESPA was formulated to have the following properties:
• A mean wind strength chosen for each simulated race, by sampling a wind
distribution. This distribution should ideally be derived from historical
meteorological data for the specific times of day and period of the tournament.
This distribution should also be adjusted for any drift in the mean wind
velocity over the period of interest. These wind velocity distributions should
include upper and lower cut-off values to handle the wind velocity limits above
and below which races are not sailed.
• Each race should model temporal variation in wind velocity and direction
based on an estimate of variance for wind velocity and direction. For example,
the wind velocity standard deviation for a 3-month period may be 8 knots,
while the standard deviation for a single race, taking only 90 minutes, may be
3 knots.
2.10 Measure of merit - summary

These factors make a strong case for the use of a full tournament model as the
measure of merit for an America’s Cup yacht design optimisation. This choice of
measure of merit has a significant impact on the design of the VPP and RMP, as well
as affecting the selection of an appropriate optimisation method.
Based on the factors outlined in the previous sections, the measure of merit for
VESPA was formulated to have the following properties:
• It should use a tournament of multiple boats, with the measure of merit being
the win/loss ratio against all competitors.
• It should include multiple races against all competitors, each race having its
own mean wind velocity.
• Wind velocity and direction for each race should vary about their mean, based
on empirically derived variance for wind velocity and direction.
· 39 ·
Chapter 3 · Constraints
3  Constraints
The primary constraints that apply to any optimisation of an ACC yacht are those
imposed by the ACC rule, which specifies significant limits on the design of both
hull and rig. These constraints need to be taken into account for each design change
that occurs during the optimisation process.
3.1 The America’s Cup Class rule

The ACC rule specifies a small number of measurements that are combined in a
simple equation, the result of which cannot exceed a given value. Although earlier
versions of the rule permitted a wide range of designs, the most recent rule, version
5, reduced this range considerably. Despite these recent amendments, there is still
considerable scope for hull shape variation within the confines of the rule. However,
this tends to be in areas other than the three key measurements of length, sail area
and displacement.
The rule also imposes significant constraints on several areas of the hull design
which must be complied with. In particular, minimum freeboard values are
specified at three points along the hull, and maximum girths are specified at the fore
and aft girth stations. In addition, the ACC rule requires that the hull shape be
convex in all directions.
· 40 ·
The ACC rule requires that the following equation be satisfied-
1.25 √ 9.8 √
24.000 (3.1)
0.686
where:
• S is the rated sail area in m2
• DSP is the displacement in m3
• L is the rated length in m
These parameters are defined by the following equations:
(1 000.1 ( 320) ) (3.2)
where SM is the measured sail area in m2
/1025 (3.3)
where W is the weight of yacht in kg.
(1 2000 ( 22.1) ) (3.4)
(3.5)
(3.6)
is the greater of -
0.3 OR 1.25 ( 2.4 ) (3.7)
is the greater of -
0.116 OR 1.8 (1/ Θ) 1 (3.8)
is the greater of -
1.6 OR 1.75 ( 1.8 ) (3.9)
is the greater of -
0 OR 1.414 1/ Φ (3.10)
4 √ 28.845 (3.11)
· 41 ·
where:
• LM is the measured length in metres
• LBG is the length between girth stations in metres
• G is the girth component of LM in metres
• FGC and AGC are the forward and aft girth corrections in metres
• FG and AG are the forward and aft chain girths in metres
• FBC and ABC are the forward and aft beam corrections in metres
• Θ and Φ are the mean of the angles port and starboard of the topsides,
measured at FGS and AGS relative to the vertical, in degrees
• FP is the freeboard penalty in metres
Figure 19 illustrates the locations at which these dimensions are measured. Note
that LBG is measured in a plane 200mm above the measurement waterline MWL.
This is an effective way of preventing designers from creating excessive overhangs
fore and aft to gain unmeasured length. If LWL was the only length measure used
this would encourage designers to utilise extreme overhangs in their designs.
Figure 19. ACC Hull Measurements
The two girth correction values, FGC and AGC, are penalties that are invoked once
FG and AG rise above certain values. In general, designers of ACC yachts aim to use
the maximum possible unpenalised girth values of 0.3m for FGC and 1.6m for
AGC, giving a total value for G of 1.9m.
A freeboard penalty, FP, is also applied for freeboards that are less than those
specified by the rule, and a weight penalty, WP, is applied if the weight is greater
than 24,000 kgs (see eq. 4.11). ACC designs usually avoid incurring these penalties
by staying within the limits specified by the ACC rule.
· 42 ·
Assuming zero values for FP and WP, equation (3.4) can be restated as –
( 1.9) 1 2000 ( 1.9) 22.1 (3.12)
This shows that if no weight, freeboard or girth penalties are incurred, L is solely
dependent on LBG and is at a minimum at an LBG of 20.182m. Deviation from
this LBG results in a rapidly increasing penalty, as shown in Figure 20.
Figure 20. Rated Length (L) versus Length Between Girths (LBG)
As can be seen from equation (3.2), rated sail area S is dependent solely on the
value of measured sail area, SM. This function is graphed in Figure 21.
Figure 21. Rated Sail Area (S) versus Measured Sail Area (SM)
To view the three-dimensional parameter space of the ACC rule, it is necessary to

restate equation (3.1) in terms of weight –
1.25 √ 24 0.686 (3.13)

W 1025
9.8
· 43 ·
Given that L and S can be expressed in terms of LBG and SM (equations (3.12)
and (3.2) respectively), W can be graphed as a function of LBG and SM, as
illustrated in Figure 22.

Figure 22. Weight (W) as a function of Length Between Girths (LBG) and Measured
Sail Area (SM)
The weight penalty WP heavily penalises displacement as it deviates from a value of

24,000 kg, leaving most of the displacement bands shown in Figure 22 unusable. In
reality, the available parameter range is even smaller, as only values in the upper
right quadrant of the design space will maximise both LBG and SM.
As a result, allowable values for LBG and SM lie on the curve shown in Figure 23.
The usable range for LBG is from 20.182m to 20.234m, a difference of 0.052m,
while the range for SM is from 318m2 to 321.4 m2, a difference of only 3.4 m2.

Figure 23. Usable ranges for Length Between Girths (LBG) and Measured Sail Area (SM)
· 44 ·
Variation in expected weather conditions for an event may result in some

optimisation of these variables taking place. For example, a light weather event
might result in designs clustering around point A in Figure 23, (i.e. moderate LBG
and maximum SM) while a heavy weather event may see designs favouring longer
LBG and lower SM values, such as at point B. However, it is clear that the
narrowness of the range for each parameter provides very little scope for significant
variation of SM and LBG design variables. For example, boats with LBG and SM
values shown by points A and B in Figure 23, would each have speed advantages
over the other of less than 0.5 seconds per mile, in the conditions that suit each best.
Constraints or bounds?
While the key dimensions of an ACC yacht are heavily constrained by the ACC rule,
these constraints are most efficiently applied at the point in the optimisation process
that design variables are modified. Rather than the optimisation process having
conventional constraints, design variations are confined to the feasible range by
adjusting dependent hull-design variables so that the ACC rule is not violated. For
example, regardless of how other hull shape parameters are varied, LBG is scaled so
that it always stays within the acceptable range, while W is kept at a value of 24,000
kg, primarily by scaling canoe body draft, TC.
As a result of this preliminary check for feasibility, the optimisation process becomes
an unconstrained problem, with simple upper and lower bounds on design
variables. Although extreme values of these bounds may result in designs that were
not geometrically achievable, in practice, given the relatively narrow ranges specified
for the bounds for each design variable, no design variable combinations used by
VESPA were found to be infeasible.
Effect of the ACC rule on design choices
Like most yacht rating rules, the ACC rule is type forming, in that it encourages
great similarity between designs due to the values chosen for trade-offs between
length, sail area and displacement, and the nature and location of the measurement
points. ACC yachts have evolved to be extremely slender, with long overhangs, deep
draft and high ballast ratios. The hulls themselves have become boxy and slab sided
with significant flat areas, joined by areas of high curvature.
A good example of these trends is illustrated by Figure 25, which shows ITA-94, the
most recent yacht of the Italian Luna Rossa team. The flat bottom and vertical
topsides are examples of the extremes to which hull shapes have developed. These
extreme characteristics have occurred, not because they are the most efficient
hydrodynamically, but because they are the best compromise within the constraints
of the rule.
· 45 ·
Figure 24. ITA-94, an example of ACC hull design extremes
There is also little scope available for the optimisation of rig parameters. The ACC
rule strictly regulates the length of the mast, together with its weight and centre of
gravity. The luff length of the mainsail P and the foretriangle height I also have set
limits. As a result, all ACC yachts have rig proportions similar to those shown in
Figure 25.
Figure 25. ACC rig configuration (© Ivo Rovira 2007, reproduced with permission)
· 46 ·
As a result of P and I values being similar for all boats, optimisation of the sailplan is
restricted to the foretriangle base value J and the horizontal mainsail girth values E1
to E5. As most teams have adopted J values of approximately 8.5m, only the
mainsail girths have been a subject of significant design variation for this Cup cycle.
For the purposes of the VESPA project, no attempt was made to optimise rig
proportions. Rather, a set of rig dimensions was adopted that corresponded very
closely to those used by Alinghi’s rig designers, and these dimensions were used for
all hull design variants. It was considered that within the range of hull shapes
examined, the standard parameters used by VPPs to vary and optimise sail forces, i.e.
reef, flat and twist, are adequate to handle the required variation in sail shapes.
Similarly, no attempt was made to optimise appendage design using VESPA, with a
single bulb, fin and rudder used for all design variations. The geometric model of
these appendages extended up into the canoe body so that decreases or increases in
canoe body draft would reveal more or less fin and rudder area. Keels and rudders
were moved fore and aft to compensate for different LCB and aft waterline ending
positions, with the LCG of the yacht moved to correspond with the LCB location and
the rudder maintaining a constant distance from the aft waterline ending.
It was considered that the design of the appendages was sufficiently independent of
the hull design variations examined, that the use of a single appendage package was
an acceptable simplification of the optimisation problem.
With the key speed producing variables of length, sail area and displacement
severely constrained by the ACC rule, there are only a small number of hull shape
variables remaining to be explored by an optimisation process. Of these, the most
obvious candidates for design optimisation are parameters controlling transverse
sectional shape (BWL, TC, CM and flare) and longitudinal area and volume
distribution (LCB, CWP, CIL and CP).
While the restricted design space makes optimisation more difficult, as the gains to
be made are likely to be small, the relatively small number of free parameters also
simplifies the optimisation significantly, with the variation of hull shape restricted to
as few as five key variables.
The narrowing of the design parameter space has also dramatically limited a team’s
ability to “re-mode” a boat between races to suit variations in weather. In 1987, the
designers of Stars & Stripes changed the ballast and sail area carried by the yacht in
order to tailor performance to the expected weather conditions for each round-robin
of the elimination series. This was possible within the 12 Metre class yachts used for
the America’s Cup at the time, which allowed a wide range of waterline lengths,
displacements and sail areas to be used. While earlier versions of the ACC rule also
allowed some re-moding to occur, the narrowing of design parameter ranges for the
version 5.0 ACC rule has prevented significant re-moding of the competing boats.
· 47 ·
Chapter 4 · Data Approximation
4  Data Approximation
In order to perform a useful optimisation of an America’s Cup Class yacht design
within a reasonable time using readily available computer hardware, it is necessary
to find ways to reduce the number of calls made to expensive analysis functions.
Although CFD methods are constantly developing and improving, they are at the
same time becoming more, rather than less, computationally expensive. Potential
flow methods, although theoretically less accurate than RANS codes, are still widely
used due to their relatively short execution times. However, potential-flow methods
can still be prohibitively time consuming when a large number of cases must be
calculated.
To make a significant reduction in the time taken to calculate the lift and drag of a
yacht being analysed, it is necessary to look at the use of an approximation derived
from a small number of sampled data points. This approximation model, known as
a surrogate, or more commonly, a metamodel, is typically calculated in advance
using samples calculated by the chosen CFD code. Alternatively, the metamodel
may be calculated in real time as required, and may be progressively refined using
additional samples as the optimisation proceeds.
Metamodels have been widely used in aerodynamic optimisation work for more
than a decade, (Greenman 1998; Simpson et al. 1998; Pierret 1999; Jin et al. 2001;
· 48 ·
Jin et al. 2002; Ong et al. 2003; Queipo et al. 2005; Zhou et al. 2007; Alonso et al.
2009). However, there are few examples of the use of metamodels in the
hydrodynamic optimisation of ships and boats, such as Duvigneau (2002), Peri and
Mandolesi (2005), Peri and Campana (2005b) and Mason et al. (2005).
In order to create a metamodel it is necessary to sample the design parameter space

to provide a set of analysis results to which the hypersurfaces of the metamodel can
be fitted. Although there are many possible sampling schemes, such as random
sampling or full factorial arrays, the approach favoured in the metamodelling
literature, summarised by Giunta and Wojtkiewicz (2003), is the use of a quasi-
random sequence having low discrepancy between the sampled points. In this case,
the discrepancy ( ) for a sequence 1, 2, 3, … with respect to the interval
, may be defined as:
(4.1)
A sequence is thus equidistributed if the discrepancy ( ) tends to zero as tends

to infinity.
The selection of a set of design parameters is not sufficient to allow CFD analysis, as
this also requires a geometric model of a complete hull surface in order to perform
its calculations. For this to occur, it is necessary to have a method that can create
from scratch a complete hull geometry meeting those parameters (parametric
modelling), or alternatively, be able to deform an existing hull design to match the
design parameters required (parametric transformation).
This chapter describes the factors that influence the choice of sampling method,
metamodel type and hull-shape representation and variation method adopted for
use by VESPA.
4.1 Data characteristics

The choice of metamodel selected for a particular optimisation task is influenced by
several factors, some of which are related to the characteristics of the source data on
which the metamodel is to be based.
In the case covered by this thesis, the data in question comes from SPLASH, a
potential flow program that has been widely used in the field of America’s Cup
yacht design since 1987. SPLASH is used by the Alinghi team for day-to-day CFD
analysis and the program was used for all hydrodynamic calculations performed in
the course of this research.
Like other CFD codes, SPLASH is a deterministic program, meaning that it will give
identical results when given identical inputs. This does not mean that there is no
· 49 ·
noise in the system. However. SPLASH may contain low levels of systematic noise
caused by round-off error, as well as discretisation artifacts caused by breaking the
hull surface into a finite number of panels.
SPLASH may also fail to converge on a solution in some cases, particularly at

combinations of high forward speed coupled with significant heel and/or yaw. These
non-converged points may have significant error associated with them, in which
case they need to be identified as outliers and eliminated. Alternatively, the error
may be small enough that the data point can be included but be treated as if it has a
small amount of random, but repeatable noise.
4.2 Metamodels
Many alternative metamodel formulations exist which may be suitable for use in the
proposed optimisation system, ranging from traditional statistical regression
through to methods that are quite recent, such as radial basis function networks.
Despite the wide range of alternatives, no method has shown to be overwhelmingly
superior across a range of approximation tasks.
Criteria used in the selection of a metamodelling method include the quality of fit
of the metamodel to the sampled data points and the smoothness of the resultant
fitted hypersurfaces. However, the choice of metamodel type for a particular
application is also dependent on several other factors:
• Quantity of data available;
• Dimensionality of the solution space;
• Complexity of the solution surface;
• Degree of noise associated with the data;
• Ease of use.
Quantity of data
For a problem where only a small number of data points are available, the
metamodelling strategy may differ from the case where a large volume of data has
been provided. For example, a neural network metamodel may be unsuitable for
small datasets, particularly as a significant proportion of the data needs to be set
aside for validation and test sets and is therefore not available for network training.
In such cases, a conventional regression model may be more appropriate.
· 50 ·
Solution complexity.
The solution space of a CFD analysis may have a few or many dimensions; the
solution surfaces may be relatively smooth or highly non-linear; and there may be a
single optimum or multiple local optima.
Some metamodelling methods, such as polynomial regression, which work

exceptionally well on smooth, unimodal solutions surfaces having few dimensions,
may be completely inadequate when applied to high-dimensional, non-linear,
multimodal solution spaces. It is essential that the characteristics of the
metamodelling method chosen is appropriate to the type of data to be fitted.
In the CFD case examined in this work, the parameter space was 10 dimensional,
the input variables for each analysis being LBG, BWL, CP, CM, LCB, heel angle, yaw
angle, rudder angle, tab angle and hull velocity. The results of Sahoo (1997),
Yasukawa (2000), Hirata (2004) and Pinto (2004) strongly suggest that solution
surfaces in this case will be both non-linear and multimodal.
Noise.
Physical experiments such as towing tank data usually include a component of

random noise which causes problems for metamodelling approaches that fit data
points exactly. For these data, it is preferable to use a least-squares regression method
that allows a smooth surface to be fitted through the noisy data.
Similarly, although some deterministic computer experiments may be completely

noise free, many CFD codes contain low levels of systematic noise caused by round-
off errors and discretisation artifacts. This type of data also benefits from a least
squares fitting approach.
Another issue is that calculations do not always converge perfectly, particularly in

cases of high angles of heel and yaw. This is illustrated in Figure 26, reproduced
from Keane and Nair (2005), which shows an objective function contaminated by
discretisation noise and occasional failed calculations.
This form of noise differs from the noise expected from physical experiments, in
that it cannot be assumed as normally distributed with a zero mean. This potential
for non-random noise occurs with SPLASH, making purely interpolative methods
inappropriate in this case.
· 51 ·
Figure 26. Objective function contaminated by noise, from Keane

and Nair (2005)
Ease of use.
Ease of use may appear to be a minor consideration. However, it encompasses

several issues that may directly affect the successful creation of the metamodel.
These include:
• Ease of implementation. Is tested and validated software available for the
metamodelling method, either source code or executable file, which allows it
to be easily implemented within the optimisation environment?
• Is the metamodelling method automatic, or does it have multiple settings
which require expert knowledge in order to tune and optimise the metamodel?
If it does have multiple tuning parameters, are the parameters and tools for
optimising the metamodel accessible and easy to use?
• Does the method provide information about the nature of the data being fitted,
or is it a “black box” method that provides no information other than the
outputs produced for a given set of input parameters?
• What measures are available to verify the quality and prediction accuracy of
the metamodel?
· 52 ·
4.2.1 Metamodel candidates
Metamodelling methods supported by the optimisation literature include

polynomial regression (PR), Response Surface Methods (RSM), (Box and Wilson
1951), Gaussian Processes, Kriging, Artificial Neural Networks (ANN), Radial Basis
Function networks (RBF), Support Vector Machines (SVM), (Burges 1998), Least
Interpolating Polynomials (De Boor and Ron 1990), and Multivariate Adaptive
Regression Splines (MARS), (Friedman 1991).
Of these, the first six methods are widely used within the field of optimisation, and
each of these has been shown to work well across a range of problems.
Several authors have performed comparative evaluations to determine the methods

that have the best performance for a range of problem domains. Jin et al. (2003)
compared polynomial regression, MARS, Kriging and RBF across a range of test
problems. Polynomial regression outperformed the other methods on small-scale,
low-order, non-linear problems. However, RBF had superior performance in all
other domains.
Simpson et al. (2000) compared a range of sampling strategies combined with four
metamodel methods: second order RSM, MARS, Kriging and RBF. The polynomial
based RSM did well approximating low-order non-linear functions, but performed
poorly with more complex problems. The least stable method was found to be
MARS, while both RBF and Kriging performed well.
Simpson et al. (2001) surveyed prior literature to provide recommendations on

metamodel selection. Simpson’s conclusions are summarised in Table 3.
Table 3. Recommendations For Model Choice And Use, After Simpson (2001)
Model Choice Characteristics/Appropriate Uses
Response Surfaces • well established and easy to use

• best suited for applications with random error
• appropriate for applications with < 10 factors
Neural Networks • good for highly non-linear or very large problems

• best suited for deterministic applications
• high computational expense
• best for repeated application
Kriging • extremely flexible but complex

• well suited for deterministic applications
• can handle applications with < 50 factors
• limited support is currently available for implementation
· 53 ·
Note that although Kriging and Gaussian Processes are both widely mentioned in
the literature, they are essentially the same thing, with the term Kriging originating
in the field of geostatistics, while Gaussian Processes is a term used by the statistics
community. Kriging/Gaussian Processes are interpolative methods, making them
ideal for deterministic computer experiments that do not have a noise component.
However, this characteristic can cause difficulty when dealing with noisy data.
When noisy data points are precisely interpolated, the result tends to be
uncontrolled oscillations in the solution surfaces, raising the possibility of false
optima being created by the metamodelling process.
The wave-making resistance of marine vessels is known to contain non-linearities.

Hirata (2004) and Yasukawa (2000) showed that even for basic monohull merchant
ships, the solution space of the hull resistance optimisation problem is multi-modal.
It is also known that CFD codes such as SPLASH may contain low level noise and
unconverged results.
As a result, it is expected that the data produced by a SPLASH analysis of a fleet of

America’s Cup Class yachts will be non-linear; with a potentially multi-modal
solution space of moderate complexity and dimensionality; derived from a
deterministic analysis but containing some element of systematic noise and non-
Gaussian random noise. Consequently, the following methods may be eliminated as
candidates, based on the recommendations contained in the references cited above:
• Polynomial regression, due to the complexity and non-linearity of the CFD
data.
• Response Surface Methods, due to the complexity and non-linearity of the
CFD data.
• Kriging is an interpolative method that fits the data exactly. Consequently, it is
not ideal for use with data containing random or systematic noise.
This leaves artificial neural networks, together with their related method, radial
basis function networks, as the preferred options for the metamodelling of high-
dimensional, non-linear, multi-modal data containing systematic noise. Regarding
neural networks, Chen et al. (2003) make the comment:
“Although ANNs are generally flexible enough to model anything, they are
computationally intensive, and a significant quantity of representative data is
required to both fit and validate the model.... However, given enough good data,
ANNs can outperform all the other previously described statistical modeling
methods.” (Chen et al. 2003, p.245)
· 54 ·
Artificial Neural Networks (ANN)
An artificial neural network (ANN) is a network composed of many simple

processors referred to as units or neurons, linked by communication channels or
connections that carry encoded numeric data. These units operate only on their
local data and input they receive via the connections, effectively applying a multiple
linear regression followed by a non-linear transformation, on the output value .
This transformation, or transfer function, is most commonly a sigmoidal function,
as shown in Figure 27. However, other transfer functions such as tangent-sigmoid or
linear may also be used.
Figure 27. Neuron with sigmoidal activation function
If the inputs to each neuron are designated , ,..., , and the regression
coefficients are designated by the weights, , ,..., , then the output, , is
given by:
1 (4.2)
1 e
where
η w x β (4.3)
and
β is the "bias value" of the neuron.
A neural network is created by assembling neurons into a network architecture.

There are many different architectures and topologies for neural networks. The
form that has found widest application in the area of metamodelling is the feed-
forward network or multi-layer perceptron (MLP), shown in Figure 28.
· 55 ·
Figure 28. Multi-layer perceptron architecture
An MLP has an input layer with a series of inputs, one or more hidden layers and an
output layer. The number of input elements is determined by the number of
variables in the input dataset, while the number of outputs is determined by the
number of result values required.
The number of hidden layers and elements in the network can vary, and finding the
optimal network architecture for fitting a given dataset is a non-trivial problem.
Rather than being programmed, a neural network “learns” from examples presented
to it based on some form of training rule. In many cases the backpropagation
algorithm (Rumelhart et al. 1986) is used to train the neural network. However,
neural networks are equally amenable to other gradient-based approaches such as
the conjugate gradient method, the quasi-Newton method or the Levenberg-
Marquardt algorithm.
According to Sarle (1994):
“MLPs are general purpose, flexible, non-linear models that, given enough hidden
neurons and enough data, can approximate virtually any function to any desired
degree of accuracy. In other words, MLPs are universal approximators.” (Sarle
1994, p. 5)
Although Sarle identifies many similarities between neural networks and traditional
polynomial regression methods, neural networks also have several distinct
advantages. Neural networks are global models, allowing them to model the entire
range of interest rather than a smaller portion. The lack of appropriate global
models has been a source of difficulty in the past for ship resistance approximation,
with vessels having different speed/length or displacement/length ratios requiring
different resistance regression models.
· 56 ·
Neural networks are non-parametric models, and consequently there is no need to

choose the functional form for each dimension, meaning that a potential source of
error is avoided. However, a non-parametric approach brings with it the risk of over
fitting to any noise or errors in the data. As a result, it is essential to use a rigorous
approach to validation of the trained neural network.
As neural networks learn by example, they do not require the traditional statistical
assumptions such as constant error variance and Gaussian distribution of errors.
Instead, the neural network user gathers representative data and invokes training
algorithms to learn the structure of the data. Although the user requires some
knowledge of how to select and prepare the data, the level of statistical expertise
required to create a useful neural network model is less than is required to create a
conventional regression model of equivalent quality.
Radial Basis Function networks (RBF)
Radial Basis Function networks (RBF) have been developed for scattered
multivariate data approximation (Dyn et al. 1986). They are similar in structure to a
multi-layer perceptron, being a feed-forward network with an input layer, hidden
layer and output layer Figure 29.
Figure 29. Radial Basis Function network architecture
Rather than use the sigmoidal activation function, RBFs use linear combinations of
a radially symmetric basis function, using a Euclidean distance, to approximate
response functions.
A radial basis function model can be expressed as:
( ) (4.4)
· 57 ·
where is a vector of inputs, are weighting coefficients, are RBF centres,

denotes the Euclidean norm, and is a nonlinear function, typically Gaussian:
exp (4.5)
RBF approximations have been shown to produce good fits to both deterministic
and stochastic response functions (Powell 1987). They have been used successfully
by several researchers in the marine optimisation field including Peri and Campana
(2005c), and Fassardi and Hochkirch (2006).
4.2.2 Metamodelling – practical usage
Both neural networks and radial basis function networks can be expected to give
good results for the lift and drag data that VESPA uses, which is derived from
SPLASH. The SPLASH data forms a solution surface, which is likely to be relatively
smooth, yet non-linear and possibly multi-modal, with the presence of some non-
random noise.
Other significant differentiating factors in the choice of a metamodel are data

management issues to do with ease of fitting, validation and visualisation. These are
primarily related to the user interface of the software involved and favour methods
that are sufficiently mature and widely used for there to be a variety of well-designed
commercial software available.
The ability to optimise the quality of fit and validate results easily is paramount, as a
small error can drive the optimisation process to give unrealistic results. Software
with good data manipulation and visualisation capabilities is therefore of great value.
4.2.3 Metamodel selection
After careful evaluation of alternatives, neural networks were selected for use in
VESPA for the creation of metamodels for lift and drag data created by SPLASH. In
previous work by the author, (Mason et al. 2005), neural networks were shown to
perform well in this type of application, using powerful commercial software
available at reasonable cost. Although RBF networks showed great promise, the lack
of available software with strong analysis and verification features prevented their
adoption.
Although the primary use of a metamodel is the reduction of the evaluation time of
an expensive function, there are additional benefits. For noisy functions, a
metamodel can smooth the data making the optimisation process simpler. For
functions that can suffer from some numerical instability, such as some CFD
programs, the process of fitting the metamodel can also be used to filter out non-
converged points for a net increase in data quality.
· 58 ·
4.3 Design of Experiments

The adoption of a metamodel as a surrogate for direct calculation of CFD data
allows for a dramatic reduction in the number of CFD calculations but requires that
the design space be sampled. As this sampling will be sparse, it is important that the
spacing of the sample points be regular so that the metamodel hypersurfaces may be
fitted without introducing excessive bias and variance.
In order to determine the most efficient way of sampling the design space, it is
necessary to refer to the field of Design of Experiments (DOE), which commenced
with the work of geneticist and statistician Sir Ronald A. Fisher, (1935). DOE is
based on several concepts originated by Fisher, including orthogonality of variables,
experiment randomisation, Analysis of Variance (ANOVA) and the use of factorial
experiments rather than the traditional one-factor-at-a-time method.
One difficulty for full factorial designs was the “curse of dimensionality”. As the
number of factors in an experiment increased, the number of experiments required
for a full factorial experimental design grows exponentially. For example, in order to
fit a linear model to an experiment having ten factors or dimensions, 210 samples are
required. If the solution surface had a quadratic form, requiring 3 points in each
dimension, a ten dimensional problem would require 310 samples.
In place of full factorial designs, DOE practitioners investigated how to minimise

the number of experiments performed while retaining the ability to accurately
determine the influence of each factor in the experiment. This resulted in various
fractional factorial designs being created, of which two of the better known designs,
the Box-Benkhen and face-centred, central composite design are illustrated in
Figure 30.
Figure 30. Full factorial, Box-Benkhen and central composite experimental designs
Early DOE work was based on the assumption that experimental results
incorporated random noise and consequently, experiments were structured to
minimise the effects of this noise. As a result, traditional DOE methods tend to have
experiments at the extremes of the parameter space to minimise linear fitting errors,
and utilise replication of experiments in order to estimate the variability of the
experimental results. Results were assumed as linear or quadratic in nature, allowing
polynomial response surfaces to be fitted to approximate the phenomenon under
investigation.
· 59 ·
With the advent of computer-based simulations, experimental design underwent

several changes. The deterministic nature of computer experiments eliminated the
need for replication of individual experiments, as replicated simulations would give
identical results, rather than being subject to random error. More complex problems
could also be investigated, with the result that quadratic approximations were no
longer sufficient. As a result, fractional factorial designs were supplanted by space
filling designs for computer based experiments, as these distributed data throughout
the parameter space, rather than primarily about its perimeter.
A simple space filling design can be created by selecting a suitable number of sample
points at random throughout the design space, known as pseudo-Monte Carlo
sampling. However, an examination of such a random sample shows clustering of
points and large areas that are poorly sampled. Random sampling displays high
discrepancy, that is, a significant deviation from a uniform distribution.
4.3.1 Quasi-random methods
One approach to achieving low discrepancy sampling is the Latin hypercube

(McKay et al. 1979), which ensures that the parameter space is sampled uniformly
by dividing the domain 0,1 along each dimension into sub-intervals, each of
which contains one sample point. In its simplest form in two dimensions, a Latin
hypercube distributes the required number of samples equally along the diagonal of
the parameter space (Figure 31d). This diagonal distribution is highly correlated and
does not sample the space in a uniform manner.
Figure 31. Quasi-random sampling methods
· 60 ·
Up to ! valid two-dimensional Latin hypercubes can be generated by permuting

the columns of the diagonal distribution. Determining which of these has the
lowest discrepancy becomes increasingly difficult as the number of samples increases,
particularly when the parameter space is extended into many dimensions.
Various methods have been proposed by many different authors for generating Latin
hypercube samplings having low discrepancy. These include; Maximin Latin
hypercubes (Johnson et al. 1990; Morris and Mitchell 1995), Minimal Integrated
Least Square Error designs (Sacks et al. 1989), Orthogonal Array based Latin
hypercube designs (Tang 1993), Integrated Mean Square Error optimal Latin
hypercubes (Park 1994) and the Uniform Design (Fang et al. 2000).
One alternative to variations on the Latin hypercube is the low-discrepancy

sequence, which allows an unlimited number of points to be added incrementally to
an n dimensional space such that discrepancy is kept low. Such sequences have been
proposed by Hammersley, Halton, Sobol, Faure, and Niederreiter, and are detailed
in Niederreiter (1992). Figure 31 illustrates several space filling experimental designs,
each having 25 sample points, including a 5 x 5 full factorial design, a random
sequence, a Latin hypercube, a Sobol low-discrepancy sequence and a Uniform
Design. It can be seen that although all sequences span the space effectively, the
Uniform Design exhibits a more regular spacing than all other methods.
Hurrion (1999) demonstrated that neural network metamodels based on

randomised sampling outperformed metamodels based on conventional regression
using full factorial sampling. Similarly, Giunta and Wojtkiewicz (2003) argued
strongly in favour of the adoption of quasi-random sampling methods for
deterministic, computer based experiments.
As a result, a quasi-random, low-discrepancy DOE method was selected for the

sampling of design parameters used to calculate the hull performance metamodel
within VESPA. The method selected, the Uniform Design, (Fang 2006), was chosen
on the basis of recommendations contained in Notz (2003) and Simpson et al.
(2000).
Although low-discrepancy sequences such as those by Sobol (1967) were considered,

the regular spacing of the Uniform Design was preferred, as the intervals could be
chosen to correspond with portions of the test matrix previously used both for CFD
analysis and for tank testing. This made comparison of neural network predictions
with existing data easier, as results from multiple sources could be compared on the
same graph. If parameter values had fallen on irregular intervals such as would occur
with a Sobol sequence, direct comparison between raw CFD data, tank data and
metamodel output would not have been possible.
· 61 ·
4.4 Hull shape representation

A yacht-design optimisation system must be able to represent different design
configurations in a numerical format, in order for these configurations to be varied,
evaluated and ranked as part of the search process. There have been many formats
used for the numerical representation of hull shapes and these can be broadly
categorised as follows:
• A wireframe representation consisting of three-dimensional curves, each
approximated by a polygon. Although this method has been widely used for
capturing hull shape data from lines drawings or actual vessels, this approach is
not currently utilised by any of the CAD systems used specifically for the
design of marine hull shapes.
• A wireframe representation consisting of three-dimensional curves defined by a
spline function such as a cubic spline, Bezier curve, B-spline or NURBS curve.
This approach was used by early ship design systems such as Autokon,
(Reenskaug 2003), and Steerbear, (McNaull 1980), which faired hand-drawn
hull lines. The method has now been superseded in systems used for ab initio
hull-design by surface based methods.
• Hybrid wireframe/surface models as used NAPA Design (Lengyel 2003). NAPA
uses a wireframe definition of sections, waterlines and buttocks as the basis for
the automatic construction of a network of surface patches.
• Recent hybrid systems such as the FRIENDSHIP Modeller, (Harries et al.
2003a), and Paramarine, (Bole and Lee 2006), use a network of parametrically
defined feature lines, such as those defining stem profile, sheerline, flat of side
and flat of bottom, to automatically generate hull surfaces.
• A three-dimensional surface representation as used by programs such as
Maxsurf, (Formation Design Systems 2006). These programs typically use a
small number of large surfaces, typically subdivision, Bezier, B-spline, or
NURBS surfaces to define a hull shape.
Of the above methods, the one most widely used by commercial marine-design
software is the NURBS surface model, adopted by interactive design programs such
as Maxsurf, Autoship, Fastship, Rhino, Prolines and ProSurf.
· 62 ·
Interactive NURBS surface modelling
The use of B-spline surfaces for defining ship hull surfaces was first described by
Rogers (1977). Non-uniform rational B-splines (NURBS), a more general form of
B-splines, came to prominence with their inclusion in the Initial Graphics Exchange
Specification (IGES), first published in 1980. NURBS were included in a
commercially available marine hull design system for the first time in 1985, being
added to Maxsurf, a program written by the author of this thesis.
A NURBS surface of degree ( , ) is defined by Piegl and Tiller (1997) as:
∑ ∑ , ( ) , ( ) , ,
( , ) 0 , 1 (4.6)
∑ ∑ , ( ) , ( ) ,
where , and , are the B-spline basis functions, , are control points, and the
weight , is the last ordinate of the homogeneous point , .
The i-th B-spline basis function of degree p, written as , ( ) , is defined

recursively by the Cox-de Boor algorithm as follows:
1 if
, ( )
0 otherwise
, ( ) , ( ) , ( ) (4.7)
A NURBS surface modeller typically presents the user with a network of control
points, as shown in Figure 32, which are moved in x, y and z directions to produce
the desired hull shape.
Figure 32. NURBS surface and corresponding control point net
Although this method allows a designer to create almost any shape that can be
imagined, the design process is manual and the achievement of a suitable hull shape
is highly dependent on the skill of the user.
· 63 ·
Alternative approaches to hull-form definition
Surface based programs have now almost completely superseded wireframe-based

programs for the ab initio design of marine hull shapes. However, the differing
requirements of the designers of large versus small vessels have resulted in the
adoption of different solutions to the question of the best process for hull design.
Typically, the designers of large commercial ships commence the design process with
a set of hull form parameters. These are derived using an initial design procedure, or
estimated from a database of existing vessels. These parameters are used to either
automatically generate a suitable hull shape, or to deform an existing design to
match the parameter values. In other words, the designers of large ships use the hull
form parameters to derive the hull geometry.
Designers of small craft typically work in the opposite direction, creating a unique
geometry and deriving the hull form parameters from it. This approach is often
required as there is a great deal more variability in the design requirements for small
craft, making it more difficult to modify a parent hull to suit. Once a hull shape has
been created, its hydrostatic properties may need further adjustment, and changing
the hull surface manually to achieve this is a trial-and-error process.
To address this problem, procedures for the parametric transformation of hull

surfaces have been created to permit a designer to automatically adjust the
hydrostatic parameters of a hull shape to meet requirements.
As a result of the development of parametric transformation functions for surface

based hull modellers, the design process for small vessels now more closely matches
the approach taken for large ships. The major remaining difference is that while the
ship designer may have the luxury of sourcing a parent design from a library of
similar existing vessels, the small craft designer will need, in most cases, to first
create a suitable parent hull shape.
Interactive hull design software now includes the ability to apply parametric
transformations to hull shapes represented as NURBS surfaces. However, during the
past ten years an alternative approach has also been developed. This approach is
typified by programs such as Paramarine, (Bole and Lee 2006), and the
FRIENDSHIP Modeller, (Harries et al. 2003a), which allow for the automatic
creation of a hull shape from a set of key parameter values and feature curves.
The development of these two different methods for creating and varying hull
shapes has had significant implications for the automated optimisation of marine
vessels. In order to choose which is most appropriate for the optimisation of ACC
yachts, it is necessary to examine possible hull variation procedures in more detail.
· 64 ·
4.5 Hull shape variation

An optimisation algorithm progresses by making small adjustments to parameter
values in order to search the available solution space. Each design variation is
evaluated and ranked against previous solutions to determine the direction in which
the next step in the solution space should be made.
There are several ways in which variation to the hull design of a marine vessel might
be achieved within an optimisation algorithm:
• direct manipulation of NURBS surface control points;
• blending of two or more parent hull shapes;
• creation of a hull shape entirely from numerical parameter values;
• deformation of a parent design using a combination of linear and non-linear
transformations.
Methods that operate on a parent hull form have an inherent advantage for ACC
yacht design optimisation, as they can preserve subtle design features that may be
difficult or impossible to capture using an approach that creates a hull shape from
scratch. Although this is not a significant problem for commercial ships, the design
of racing yachts can involve subtle adjustments that may be crucial to the final
performance of the yacht. The precise capturing of a designer’s intent at this level of
detail may be beyond the capability of a parametric modelling system, or may
require so many interrelated parameters that optimisation becomes unwieldy.
Conversely, a parent hull captures this design intent precisely, and this can be
retained so long as the procedure used for hull variation is well designed.
4.5.1 Direct control point manipulation
Although direct control point manipulation is widely used by interactive CAD

systems for ab initio hull design, it is not ideal as a strategy for shape control in an
automated optimisation procedure, due to the large number of variables required.
For example, Hendrix et al. (2001) performed an optimisation of the resistance of a
Wigley hull using a 5 × 9 control point mesh, each with x, y, z co-ordinates. By
constraining the movement of many of the control points, the available degrees of
freedom were reduced from 135 (i.e. 5 × 9 × 3) to 61. No constraints for fairness
and convexity were enforced.
While the optimisation of a simple form like a Wigley hull may be feasible using
this approach, an ACC hull is more difficult. A typical ACC hull might be modelled
with a control point mesh of 10 columns and 8 rows, giving 240 degrees of freedom,
with a significant number of constraints required. In particular, fairness and
convexity constraints require that movement of adjacent control points be correlated
· 65 ·
to some extent, as the movement of a single control point would be likely to

introduce a bump or hollow in the hull surface.
The large number of parameters and constraints, and the requirement that
movements to adjacent control points be correlated, makes successful optimisation
of ACC hulls using this approach difficult if not impractical.
4.5.2 Blending
Where multiple parent models exist having characteristics that may be appropriate
to the final design, it may be possible to perform a weighted blend of two or more
parent hulls to create a new design.
Neu (2000a; 2000b; 2007) used a barycentric blend of the form:
Resultant Ship Hull Basis Hull (4.8)
where,
Blending Coefficient for Basis Hull (4.9)
1 (4.10)
and,
0 1, (4.11)
1, 2, … , N (4.12)
To perform a barycentric blend, the parent hull shapes must use the same numerical
representation, which requires a method for mapping equivalent parameterisations
for each hull surface. This limitation does not occur if the hull surfaces being
blended are NURBS surfaces, having control point nets of equivalent degree and
dimensions. In this case the blending occurs, not between points on the surface, but
between vertices in the control point net.
This is illustrated in Figure 33, which shows the effect of blending two bow sections
from different ACC hulls. Parent hull A has a form referred to as a Davidson bow, as
used by Team New Zealand to win the 2000 America’s Cup. This combines a
knuckle below the waterline with a steeply upswept forward overhang, designed to
maximise sailing length for a given LBG. Parent B uses a “destroyer” style bow with
shorter overhang and no knuckle.
· 66 ·
Figure 33. Blending of ACC bow sections
These two shapes are blended together in a 60:40 ratio to obtain the child hull
shape. In this case, each control point for the child hull surface is calculated as;
Child CP , 0.6 ParentA CP , 0.4 ParentB CP , (4.13)
The blending approach allows existing hulls with known qualities to be used as
parents. If the parent hulls are fair and the mapping of equivalent points on each
hull is good, derived hull shapes should also be fair. However, the barycentric
blending method suffers from the problem that variations to hull parameters are not
independent. The blending coefficient Cn applies to each hull in its entirety, and as a
result hydrostatic parameters of the blended hull are fully correlated to one another.
As an example, assume two parents, one with both high CP and high CM, and one
with low CP and low CM. In this case, there is no way to blend the two parents to
achieve a child hull that combines a high CP with a low CM, nor is it possible to
achieve a hull form that has a low CP with a high CM.
This is a significant limitation for the use of the blending approach as a variation
method within an optimisation system, as it prevents large portions of the design
parameter space from being explored.
Although the usefulness of blending as the sole variation method is limited, the
approach does have benefits when combined with other methods such as parametric
transformations. This is particularly true when used within an optimisation system
based on an evolutionary algorithm, where the blending method may be used as a
form of recombination operator (see Section 6.4.2, Recombination), allowing
variations on geometric features that are independent of the hull form’s dimensional
and hydrostatic parameters, to be explored effectively.
· 67 ·
4.5.3 Parametric modelling
Parametric modelling, as elaborated in Harries and Abt (1999b) and Harries et al.
(2001a), defines hull shapes using a set of key parameters and constraints. These are
satisfied simultaneously by a search algorithm, which gives as its output a hull shape
that matches all the required values.
The parameters specified may be overall dimensions, such as LWL, BWL, Tc; they may
be hydrostatic values, such as LWL, BWL, Tc, ∇, LCB, CM, CP and CB; or they may be
positions, slopes and curvatures of individual defining curves, such as a stem profile
or sheerline. An example of the parameters that might be specified for the design of
an ACC hull is shown in Table 4. In this case, an ACC hull was created within the
FRIENDSHIP parametric modeller using 44 parameters.
Table 4. Parameters Used to Describe an ACC Hull, From Harries et al. (2001)
· 68 ·
Parametric modelling is an ideal approach to use on simple geometric shapes. In the

case of the optimisation of semi-submersible oil rigs, Birk (2005) was able to define
the geometry of the semi-submersible pontoons and struts with a small number of
parameters. For commercial ships that have flat-of-bottom and flat-of-side joined by
a bilge radius, the number of parameters increases. However, the ship can still be
defined with a relatively small number of variables.
ACC hull shapes, on the other hand, have specific geometric features, which are
consequences of the locations at which the hull is measured. The resulting hull
shapes have bumps and flats in unusual places to gain the greatest benefit under the
ACC rule, and the specific form of these features is subject to the preferences of the
individual designer.
These geometric peculiarities make it difficult for a parametric modeller to form

hull shapes that meet the requirements of the designers involved in ACC design. As
an example, an ACC hull shape resulting from an optimisation study performed by
Harries ((2001a)) using the FRIENDSHIP Modeller is shown in Figure 34.
Figure 34. An ACC canoe body optimised using FRIENDSHIP, from Harries et al. (2001)
Although the parametrically modelled hull is fair and, accordingly, is likely to be

hydrodynamically efficient, the hull bears little resemblance to actual ACC hulls.
The plan and profile view of Harries’ hull shape is compared with a typical 2007
ACC hull in Figure 35. Note that the hulls in this illustration have been compressed
along the longitudinal axis by 50% in order to emphasise shape differences. It is
clear from the angular nature of the modern ACC hull that there are many design
features that have not been captured by the parametric modeller.
Figure 35. Comparison of parametrically modelled ACC hull with typical 2007 ACC design
· 69 ·
Baik and Gonella (2005) also used the FRIENDSHIP modeller to optimise ACC
hulls, choosing to specify the hull shape using 51 parameters. Once again, hulls
produced by the FRIENDSHIP parametric modeller, although showing more
realistic cross-sectional shapes than achieved by Harries, failed to capture significant
design features.
In addition, Baik and Gonella failed to properly constrain for hull surface convexity,
a requirement under the ACC rule, as well as neglecting to constrain the hull
dimensions to satisfy the fore and aft girth requirements. This resulted in unrealistic
hull forms, as shown in Figure 36.
Figure 36. Cross sectional shapes of optimal hulls, from Baik and Gonella (2005)
Hull A, the optimal shape found using SHIPFLOW for the zero heel case, has a very
narrow beam and a particularly narrow transom, indicating that its aft girth
measurement would be less than the maximum permitted. Hull B, the optimum
found using SHIPFLOW for the 10° heel case, has an extremely wide transom and
large aft girth measurement, and would clearly not measure as a legal ACC hull.
It is likely that the parametric modelling of a realistic ACC hull shape, although not
out of the question, would take substantially more than 50 parameters in order for
the shapes produced to be acceptable to current yacht designers involved with the
class. The use of such a large number of parameters and constraints increases the
complexity of the optimisation process and result in slower analysis and convergence.
Although parametric modelling is a promising technology, particularly in the realm

of commercial ship design, the complexity of the method and the difficulty of
achieving hull shapes that are acceptable at the highest levels of ACC yacht design
resulted in it being an unattractive option for this research work.
· 70 ·
4.5.4 Parametric transformation
Simple linear transformations, such as the scaling of length, beam and depth can be
applied to a hull shape. However, this is seldom sufficient to achieve the variation
required for optimisation. To achieve variation of hydrostatic parameters such as LCB,
CB, CP and CM, it is necessary to apply non-linear transformations to the hull
geometry.
Several schemes have been proposed to perform these non-linear transformations.

One of the first descriptions of a technique to parametrically modify the LCB and CP
of an existing lines plan was provided by Lackenby (1950). Lackenby’s method
consisted of a technique for moving sections of the hull forward and aft according
to a quadratic function to match prescribed parallel mid-body length and location
as well as LCB and CP values. However, other parameters such as ∇, water-plane area
and LCF could vary in an uncontrolled manner.
Lackenby’s method did not attempt to modify the sectional shape of the vessel, with
CM remaining constant and changes to CB limited to what could be achieved
through longitudinal changes to the volume. Although variations on Lackenby’s
method have been proposed by various authors, (Volker 1954; Söding and Rabien
1977), these refinements did not address the lack of control over transverse
hydrostatic properties, limiting the usefulness of the approach for ship hull
optimisation procedures.
The introduction of NURBS surface modelling systems for the definition of hull
shapes introduced an additional difficulty. Lackenby’s method operates directly on
hull sections, with intermediate sections needing to be interpolated and faired. This
was a straightforward process when the hull definition was stored as a lines plan or
as a three-dimensional wireframe. However, a NURBS surface is defined by the
locations of its control points and its hull sections are calculated as required from
this surface definition.
For NURBS surfaces it is not possible to move hull sections directly; rather, it is the
control points for the surface that need to be adjusted. In order for a transformation
such as Lackenby’s to be applied to NURBS surface models, it is necessary for the
surface control points to be repositioned by the transformation.
Markov and Suzuki (2001) addressed this issue, moving columns of control points
based on the output of a Davidson-Fletcher-Powell optimisation algorithm. This
approach is not necessarily fairness preserving, whereas the suggestion of Halley
(1987) that the transformation be based on a smooth piecewise cubic polynomial in
the form of a cubic spline results in fair deformations of the parent hull.
· 71 ·
Free form deformation
One potential solution to the problems inherent in Lackenby’s method is the use of
a parametric transformation procedure based on the Free Form Deformation (FFD)
of Sederberg and Parry (1986).
FFDs have been widely used by the computer graphics and animation industry to
permit objects to be deformed in a fluid manner, but have only recently been
adapted for use in design and optimisation. Peri and Campana (2005a) successfully
used FFD transformations of a cruise-ship hull form in order to optimise its sea-
keeping qualities. Menzel et al. (2005) investigated the use of FFDs in the
evolutionary optimisation of turbine blade aerofoils, finding that they resulted in a
less complex genome and improved optimisation performance.
FFD details
An FFD consists of a trivariate Bernstein polynomial, such as a Bezier, B-spline or

NURBS volume, in which the shape that is to be deformed is embedded. Deforming
this volume by moving one or more control points results in a corresponding
change to the embedded shape.
To create an FFD as defined by Sederberg and Parry, a local co-ordinate system is

imposed on a local parallelepiped region, known as a lattice space.
A point X within the lattice space has ( , , ) coordinates in this system such that:
X X S T U (4.14)
where
0 1, 0 1, 0 1
The ( , , ) coordinates of X can be found using linear algebra. A vector solution is:
T U · (X X ) S U · (X X ) S T · (X X )
, , (4.15)
T U·S S U·T S T·U
A grid of control points , , is imposed on the parallelepiped, forming 1 planes

in the S direction, 1 planes in the T direction, and 1 planes in the U
direction.
X (4.16)
where
0, … , ; 0, … , ; 0, … , ;
· 72 ·
An example of a tessellated sphere enclosed by a lattice space defined by S, T, U

directions, with its origin at X , and having an 2, 2, 2 control-point
grid, is shown in Figure 37.
Figure 37. FFD lattice enclosing a tessellated sphere
A deformation to the embedded shape is performed by moving a lattice control

point from its undisplaced position in the lattice. In this case the deformation
function is defined by a trivariate tensor-product Bezier volume, (Bezier 1974). An
arbitrary point within the lattice space is found by first calculating its
( , , ) coordinates using equation 4.15, with the deformed position being
calculated using:
( ) ( ) (u) (4.17)
where is a vector containing the Cartesian coordinates of the displaced point,

is a vector containing the Cartesian coordinates of the control point, and is a
Bernstein polynomial of the form:
(1 ) (4.18)
and
! (4.19)
!( )!
Such a deformation is shown in Figure 38, which shows the displacement of control
point , , from the FFD lattice, with the resulting deformation of a sphere
embedded in the lattice space.
· 73 ·
Figure 38. Deformation of an FFD lattice
Sederberg’s definition of an FFD utilised a tensor product volume based on the

Bezier formulation. However, an FFD can also be created using B-splines or NURBS.
Using NURBS provides an additional element of control over the Bezier and B-
spline formulations, as it is possible to deform the embedded object by changing the
weight of each control point, rather than modifying the control point’s , ,
coordinates.
In this case, the deformed position for an arbitrary point within the lattice
space, given its ( , , ) coordinates, can be found using:
∑ ∑ ∑ , ( ) , ( ) , ( ) , , , ,
(4.20)
∑ ∑ ∑ , ( ) , ( ) , ( ) , ,
where , , are control points, the weight , , is the last ordinate of the
homogeneous point , , , and where , ( ), , ( ) and , ( ) are the B-spline
basis functions of degree , and .
Modifying the control point weight values has several advantages in the context of
the parametric transformation of marine hull surfaces, as these typically require the
preservation of fairness and tangency rather than the introduction of geometric
features that would compromise the hydrodynamic properties of the vessel.
A simplified example is shown in Figure 39, where a hull section, in this case a
circular arc, is shown mapped into the unit square. A conventional FFD defined by
a NURBS bicubic patch is shown top right, with its control points adjusted to bring
about a deformation of the patch along with its embedded curve. This FFD, while
providing quite powerful shape modification capabilities, has 32 degrees of freedom
(i.e. x and y coordinates for 16 control points).
· 74 ·
Figure 39. Free Form Deformation of an embedded curve
An alternative to the conventional FFD is shown in Figure 39 bottom left. In this

case a bilinear NURBS patch is defined as the FFD surface and 3 of its 4 control
points have their weight values increased. As the control-point weight values along
the bottom and right edge of the surface are increased, the parametric spacing of the
surface compresses in the direction of the increased weights, and the embedded
curve distorts in a fair manner.
If the embedded curve in this example is regarded as representing the midship

section of a yacht, the combination of weight change with deformation of a linear
FFD achieves effective variation of BWL, T, CM and topside flare, while automatically
constraining flat of floor, convexity, fairness and tumblehome.
The effect of the increase in weight values in Figure 39 bottom left is to increase the
CM of the section without changing the tangent direction of the section at the hull
centreline and sheerline.
By moving one FFD control point in the x direction, as shown in Figure 39 bottom
right, a measure of topside flare can also be introduced. These examples show that it
is possible to make functional changes to a hull section by changing only 4
parameter values, while maintaining fairness, convexity, tumblehome and flat of
floor constraints.
· 75 ·
The use of weight changes and small control point movements within an FFD can
provide excellent control over the longitudinal form parameters of a hull design.
Several examples are illustrated in Figure 40, which shows a half-plan view of a
yacht’s sheerline and waterline undergoing various deformations by an FFD.
Figure 40. Effects of FFD weight and control point changes on

longitudinal volume distribution
· 76 ·
Figure 40 (a) shows the original hull lines overlaid onto an undistorted FFD. Figure
40 (b) and (c) show how increasing or decreasing the weight values at each end of
the FFD changes the volume distribution, effectively increasing or decreasing the CP,
CWP and CIL values for the hull.
Reducing weight values at the aft end of the hull (d) results in a shift of LCB and LCF
forward, as does forward movement of the centre column of FFD lattice control
points (e). Analogous shifts aft of LCB and LCF are shown in (f ) and (g). Note that
the weight changes in (d) and (f ) have very similar effects on the hull lines when
compared to the control point shifts of (e) and (g) respectively.
The effect of small transverse movements of FFD control points is shown in (h) and
(i). This direct control over the fineness of the ends of the hull is directly applicable
to the design of ACC yachts, due to the restrictions in the ACC rule regarding
maximum girth limits at the fore and aft girth stations.
4.5.5 VESPA hull shape transformation
Based on the strengths and weaknesses of the methods investigated, a hierarchical

FFD based parametric transformation method was chosen to perform the hull
variation required. One of the most important advantages of such a system is that a
known parent hull, encapsulating the designer’s knowledge and preferences, is used
as the basis for any design variation.
An existing NURBS surface representation of an ACC design contains an enormous

amount of information, sometimes the result of years of research and millions of
dollars in development costs, and it is important that design features be preserved
wherever possible. If an approach that involved creating each hull solely from
parameter values were to be adopted, it would not necessarily be easy to reproduce
the parent hull design’s features to the precision required.
It is important that transformed hulls do not sacrifice fairness, convexity or flatness

where required. As the applied transformations are curvature continuous, they are
fairness preserving. Constraints are also applied to the transformations to ensure
that the result also preserves convexity or planarity requirements. These are
determined by examining the sign of the principal curvatures at multiple points on
the NURBS surface.
Hull shapes resulting from the transformation must also meet the requirements of
the ACC rule. These include the measurement of length in a plane 200mm above
the water plane, a total displacement of 24,000 kg, a limit on girth values at the
FGS and AGS stations, and a requirement that the hull be convex in all directions.
· 77 ·
Finally, it is important that the transformation method used be able to function

using a small number of design variables. Such parsimony in the description of
transformations to a hull shape is of great importance to a procedure used for design
optimisation, as the number of parameters used to describe variations in hull design,
directly affect the speed and efficiency of the optimisation process.
Hierarchical FFDs
Although a global 3D FFD lattice gives good control of longitudinal hull form
parameters such as LCB, LCF, CP and CWP, it does not necessarily give precise control
of the transverse sections of the hull. One problem is illustrated in Figure 41, where
a section shape is distorted by changing the weights of the global FFD. Rather than
the section being tightly bounded by the limits of the FFD, as was the case in the
examples shown in Figure 39, the curve in this example is significantly smaller than
the bounding FFD. In this case, although the sectional shape is deformed in the
correct way to increase the CM of the hull, the vertical location of the section is
affected, as is the vertical and horizontal scale.
Figure 41. Deformation of hull section using a global FFD
These unwanted side effects have a significant effect on the ability of the FFD to
perform meaningful changes to the hull shape, as changes to one parameter, such as
CM, would have an effect on the draft and beam of the section, as well as the shape
and location of the hull centreline profile and sheerline.
In order to avoid this problem, a hierarchical FFD has been created in order to
provide precise control over hull sectional shape. In this method, the global 3D FFD
lattice encloses a series of local planar FFD lattices that lie in the planes of the
transverse columns of the hull surface control-points. Each column of control
points is projected into its own local FFD. Changes to the shape of these control
point columns are dependent on the deformations applied to the local FFDs by the
global FFD.
· 78 ·
As each column of control points is mapped onto the unit square, changes to the
column are constrained to this space. This means that the vertical and horizontal
dimensions of the embedded control point column do not change during a local
FFD transformation based on variation of the local FFD control point weights.
An ACC parent hull surface, with its enclosing global FFD and a series of planar
FFDs located at the position of each hull surface control-point column is shown in
Figure 42, together with a transformed version of the parent design. Changes have
been made to CP, LCB, BWL, CM and flare, while constraining AG and FG.
Figure 42. Transformation of an ACC hull using a hierarchical FFD
Variation of hull parameters can occur in several ways using this approach:
• Changes to the length, beam and depth of the hull are applied via scaling of
the global 3D FFD lattice.
• Changes to longitudinal volume distribution are achieved by adjusting the
weights of lattice control points in unison to deform smoothly the FFD and its
embedded surfaces. This can be seen in Figure 42, which shows a hull shape
before and after parametric transformation using an FFD. Changes to the
· 79 ·
longitudinal volume distribution in this case include a reduction in CP and an

aftward shift of LCB and LCF.
• Changes to the fore and aft taper of the hull can be adjusted by moving the
outer control points at each end of the global FFD in a transverse direction.
This feature is useful for ACC design due to the constraint on girth length at
the fore and aft girth stations. To achieve maximum girths it is necessary to be
able to adjust the beam of the hull independently at the ends relative to the
middle.
This feature also allows some control of LCF independently of LCB position. If
the aft end of the global FFD is made wider while the control point weights at
the aft end of the global FFD are reduced by a suitable amount, the LCF can be
moved aft, while the LCB is held stationary. If these variations are changed
appropriately, LCF can be held stationary while LCB is moved forward.
• Changes to the fullness of each section are achieved by modifying the control-
point weight values of each local 2D FFD as illustrated in Figure 39. The
original hull in Figure 42 has a high CM, which is reduced somewhat in the
transformed hull by reducing the weight values on the outer and lower edges
of each planar sectional FFD.
• Changes to hull flare are achieved by vertical tapering the global FFD so that
the top of the lattice is wider than the bottom in the areas where flare is to be
introduced. This in turn modifies the shape of each of the embedded 2D FFDs,
tapering each one by different degrees based on its longitudinal location in the
global FFD. This is also illustrated in Figure 42, where the vertical topsides of
the upper hull are flared in the transformed hull.
Note that the degree of flare throughout the hull is controlled by the
movements of only one or two control points in the global FFD. Flare can be
introduced amidships and tapered out fore and aft, or it can have its maximum
at the transom and reduce forward, but in each case the changes are smooth
and occur automatically.
4.6 Summary
Following a review of the alternatives, the methods adopted for use in VESPA
included:
• sampling based on the Uniform Design;
• hull geometry representation using NURBS surfaces;
• a parametric transformation of hull shapes based on a hierarchical FFD;
• a neural network based metamodel.
· 80 ·
Chapter 5 · Performance Evaluation
5  Performance Evaluation
Chapter 4 described the way in which hydrodynamic data are sampled and analysed
for a variety of hull shapes, with the results encapsulated for rapid retrieval in a
neural network based metamodel for use by VESPA.
When combined with an aerodynamic model for an ACC sail plan, this
hydrodynamic data may be used to calculate actual sailing performance for any
yacht design generated during the exploration of the design space. This creation of
sailing performance data is the role of the Velocity Prediction Program or VPP.
Once VPP data can be produced for multiple yachts, a Race Modelling Program, or
RMP, may be used to perform a simulation of one or more boats sailing a specific
race course, having a set number and location of turning marks, in a stochastically
derived wind field. Boats may be raced against one another multiple times using the
RMP, with the results tallied to give a probability of each boat winning against a
fleet of others.
The design of the RMP is crucial to the success of the overall optimisation system.
The RMP must include an appropriate level of detail in its simulation and should
not expend valuable time simulating features that are not relevant to the
determination of yacht performance.
· 81 ·
5.1 VPP
A Velocity Prediction Program (VPP) calculates the performance of a yacht when
sailing. Modern VPPs are primarily based on the work performed in the Pratt
project at MIT reported by Kerwin (1978). VPPs have undergone widespread
development due primarily to their incorporation in the International Measurement
System rule (IMS 2004) for racing yachts. Improvements have been proposed by
various authors (Letcher 1974; Milgram 1993; Schlageter and Teeters 1993; van
Oossanen 1993; Oliver and Claughton 1995; Claughton and Oliver 2003).
Given true wind direction βTW and wind velocity VTW, a VPP balances aerodynamic
forces: lift, drag and heeling moment, against the hydrodynamic forces, lift (from
the keel), resistance and righting moment, in order to determine the equilibrium
velocity for the yacht. Some VPPs also balance aerodynamic yaw moment against
the corrective moment generated by the lift produced by the yacht’s rudder.
Forces acting on a yacht sailing upwind are shown in Figure 43. Note that the
aerodynamic thrust vector A and the hydrodynamic resistance vector H are equal
and opposite, indication that the hull velocity is at equilibrium. Similarly, the
transverse component of the aerodynamic sideforce vector A and the
hydrodynamic lift vector H are in balance.
Figure 43. Forces acting on a sailing yacht
In addition, heeling moment from both the aerodynamic and hydrodynamic lift
vectors must be balanced by yacht’s righting moment. This righting moment is the
displacement of the vessel multiplied by the lever arm between the vessel’s heeled
transverse centre of gravity and its transverse centre of buoyancy.
· 82 ·
Inputs to a typical VPP include either hull lines, which are processed to determine
key dimensions, or direct input of those dimensions in tabular form. Additionally, a
VPP requires information regarding hull stability, mast and sail dimensions, as well
as details of appendage sizes and shapes. Outputs from a VPP for a given βTW and
VTW typically includes hull velocity VB, apparent wind angle βAW, apparent wind
velocity VAW, heel angle, heeling moment and righting moment. A VPP may also
provide a value for leeway angle λ.
A VPP may also provide information about the adjustments made to the sail force
model in order to balance the aerodynamic and hydrodynamic forces. The three
values commonly used by VPPs for adjusting sail configuration in order to optimise
boat speed for a given point of sail:
• Flat, a value that varies from 0.0 to 1.0 and is analogous to the degree of
reduction in the camber of the sails and therefore the amount of lift produced.
Flat tends to be applied at higher wind velocities for βTW < 60°.
• Reef, a value that varies from 0.0 to 1.0 and is equivalent to a reduction in luff
length of the mainsail. The reef parameter is used to lower the centre of effort
of the sails in order to reduce heeling moment. Reef tends to be applied at
higher wind velocities for βTW >60°. For ACC yachts, the reef parameter is not
used to change mainsail parameters, as mainsails are always used at full hoist,
although it may be used to reduce headsail area by reducing foot length.
• Twist, analogous to allowing the head of a sail to twist to reduce power.
Permits a reduction in heeling moment by changing the spanwise lift
distribution of the sails. This is of value when a rig needs to be depowered
without lowering the aspect ratio of the rig or reducing the lift coefficient of
the lower portion of the sails.
In the case of VESPA, the VPP used was provided by the Alinghi design team.
Named PAP and written by Manolo Ruiz de Elvira of Nautatec, this VPP has been
developed over a period of 10 years and was first used by the Bravo España
Challenge for America’s Cup 2000. PAP operates with two degrees of freedom,
finding an equilibrium solution by balancing forward thrust against drag, and
heeling moment against righting moment. No yaw balance is performed by PAP.
Although it is possible to incorporate a calculation of added resistance in waves

within the VPP, this has not been included for use in PAP at this time, as the small
amplitude and high frequency of the seas off Valencia does not create sufficient
variation in added resistance to justify the additional computational effort.
In order for PAP to be integrated into the VESPA optimisation system,

modifications were required to allow it to use neural-network metamodels for
hydrodynamic data. In addition, PAP was modified to allow it to be called as a
dynamically linked library (DLL) component within the VESPA system.
· 83 ·
5.2 VESPA Race Modelling Program

Prior to reviewing the choices made in the design of the VESPA Race Modelling
Program, it is worthwhile describing some of the many possible interactions
between boats, each of which may influence the outcome of a yacht race.
5.2.1 Match racing tactics
Interactions between yachts may take the form of right-of-way situations, where one
yacht must keep clear of another, losing ground in the process. A yacht upwind of
another also has the ability to disturb the flow of air encountered by another yacht’s
sails, reducing the performance of the leeward yacht. Alternatively, the threat of
disturbed air may result in an affected yacht choosing a less favourable course.
In order to simulate a match race with sufficient accuracy it is necessary to

understand these interactions and the significance of their effects on the race
outcome. In some cases, the effects are minor and are not relevant to the
determination of the faster design. In other cases, the effects are important but may
be approximated in a less complex manner than performing a detailed simulation.
The following list of interactions between boats is by no means exhaustive, and does
not include the myriad tactical manoeuvres that make up a match-racing
helmsman’s arsenal. However, the situations described are some of the most
significant in their effect on the outcome of typical match races.
Wind shadow and backwind
When yachts sail in close proximity, each boat disturbs the wind field that it passes
through, potentially affecting the wind encountered by other yachts. This is
Figure 44. Wind shadow effect of leading boat, A
· 84 ·
Boat A creates a wind shadow that extends to leeward in the direction of the
apparent wind, as well as behind and slightly to windward of the yacht’s centreline.
This wind shadow is made up of two components; turbulence, which extends to

leeward of the yacht for a distance that is dependent on the height of the rig; and
upwash, the local distortion of the wind field flowing over the yacht’s sails.
Yacht B is directly affected by the wind shadow extending downwind from yacht A.
Yacht C, while not affected by turbulence created by the rig of yacht A, is affected
by its upwash, seen as a small veer in the wind direction. This effect, commonly
referred to as backwind or lee-bow effect, results in yacht C not being able to point
as high as yacht A, resulting in a loss of ground.
Safe leeward
Although disturbed wind can affect a yacht behind and to leeward, two yachts may
be able to sail close and parallel windward courses where their wind shadows do not
affect one another. This is referred to as the safe leeward position, as shown in
Figure 45. Each yacht is vulnerable to the other moving slightly ahead, as this will
result in the slower yacht encountering disturbed air. However, if the two yachts are
of similar speed, they may be able to maintain such a close parallel course for some
time without disadvantage to either.
Figure 45. Safe leeward position
Covering
The leading boat may use its wind shadow to adversely affect the performance of a
trailing boat. In Figure 46, yacht A crosses in front of yacht B and tacks so that her
wind shadow falls directly onto B’s sails. In practice, this is such a damaging
position for B that she will invariably tack immediately to obtain clear air.
· 85 ·
Figure 46. A tacking into a tight cover on B
The lee-bow
When two boats sailing upwind on opposite tacks converge, the boat on starboard
tack has right of way. Despite this, the port tack boat may have tactical control of
the situation if it is level with or slightly ahead of the starboard tack yacht. Figure 47
illustrates this situation. At position 1, yacht A is on starboard tack and has right of
way. Yacht B must choose to tack or cross behind A. However, if yacht B tacks into
a lee-bow position, as shown in position 2, yacht A will be backwinded and will lose
distance. In this situation the best option for yacht A is to tack immediately to gain
clear air, as shown in position 3.
Figure 47. Tactical control of A by port tack yacht, B
· 86 ·
Laylines
Once a trailing yacht reaches the layline, a choice must be made as to whether to
tack on the layline behind the leading boat, or to continue one to two boat-lengths
before tacking, in order to ensure clear air. These options are illustrated in Figure 48.
Figure 48. Trailing boat options at the layline
If the boats are close and yacht B tacks line astern of yacht A, as illustrated by path 1
in Figure 48, the effect of the backwind from yacht A will be significant and yacht B
will rapidly lose one to two boat-lengths. If yacht B continues on starboard tack
slightly further prior to tacking, path 2, clear air will be guaranteed, however the
extra distance sailed is distance lost to the leader.
Which of these two options results in the smallest loss is determined by the
proximity of the yachts to the mark. When the leading yacht is close to the mark,
the period of time during which the trailing yacht is backwinded is brief, limiting
the loss. If the yachts have a significant distance to sail before the mark is reached,
sacrificing some distance initially to gain clear air will result in a smaller overall loss.
This option is important in the consideration of the disadvantage suffered by

trailing boats within a race simulation. Although the trailing boat is disadvantaged
once the layline is reached, the magnitude of that disadvantage can always be
limited to the extra distance that must be sailed beyond the layline in order to
obtain clear air.
· 87 ·
Mark rounding
The effect of the trailing boat disadvantage at each windward layline is that it is
extremely difficult for a trailing boat to be overlapped with the leading boat as the
windward mark is rounded.
This is not the case at leeward marks, where yachts have the option of choosing to
round either end of a gate, allowing them round level with another boat without
loss. The gate also allows a trailing boat to gain clear air and separation from the
leading boat, as shown in Figure 49.
Figure 49. Leeward gate options for trailing boat
Start line advantage
A yacht that is able to obtain the starboard end of the start line at the
commencement of a race has an inherent advantage due to the right-of-way status
this confers. Any tournament simulation needs to ensure that competitors are
allocated the favourable starting position in an equal number of races.
Figure 50. Right-of-way advantage at the start
· 88 ·
5.2.2 RMP background
Race Modelling Programs came to prominence with their use by the Sail America
Team for the design of Stars & Stripes 87, the winning yacht in the 1987 America’s
Cup (Oliver et al. 1987). Other RMPs have been developed since that time, most
notably the ACROBAT RMP, described by Philpott et al. (2003; 2004). These RMPs
have ranged from simple probabilistic models through to full time-domain
simulations, with execution times varying widely according to the complexity of the
simulation performed.
Sail America
Sail America used two different RMP, a simple probabilistic model, RMP1, and a
time-domain race simulation, RMP2.
RMP1
RMP1 focussed on evaluating the probability of a particular yacht winning a race

against another yacht. It was based on work described by Letcher (1974) and
assumed a wind velocity that would vary throughout the race according to a
probability density function ( TW ).
Where the relative performance of two yachts, A and B, can be displayed as a

comparison plot in which the time difference curves cross only once, the probability
that yacht B beats yacht A can be expressed as ( TW ), where TW is the crossover
velocity.
( TW ) is the cumulative probability density distribution:

TW
( TW ) ( TW ) TW (5.1)
Letcher recognised the influence of random variations in conditions on the outcome

of races and attempted to model this using an “uncertainty function” (∆ ; TW )
which gives the probability (win) that B wins over A when her computed time
difference is greater than ∆ :
0 ∆
(win) (∆ ; TW ) 0.5 0.5(∆ / ) – ∆ (5.2)
1 ∆
where 800/ TW , a subjective measure of the margin required to ensure a win.

This value was derived from the opinions of sailors as to what constituted an
unbeatable lead in a race.
· 89 ·
The probability that B wins a single race is:
(win) ( TW ) ∆ ; TW TW (5.3)
RMP1 integrated (5.3) using a ( TW )

derived from historical wind velocity
distributions from the regatta site of interest. While RMP1 had the advantage of
extremely short run times due to its simple structure, the use of the subjective
parameter to quantify uncertainty in the race results was a questionable approach.
RMP2
The second RMP developed for Sail America, RMP2, was a time domain simulation
that raced two boats around a full America’s Cup course, using wind distributions
derived from wind data for the previous 12 years for Fremantle, the site of the 1987
America’s Cup. RMP2 implemented interactions between boats, as well as penalties
for various disadvantageous tactical situations, including mark roundings. An
uncertainty function was applied to finishing times to determine a win/loss
probability for each race. Race simulations were repeated for each day of the months
of October through January for the previous 12 years, totalling almost 1500 runs.
RMP2 was used for the majority of the race modelling performed by Sail America.
The number of runs required was relatively small, as independent variation of
individual design parameters was not attempted. Rather, the comparison was
limited to allometric variation of length, constrained by the 12-Metre rule and
scantling requirements. The output from this set of tests, expressed as winning
probabilities, showed surprising complexity, as illustrated in Figure 51.
Figure 51. Win probabilities for various Stars & Stripes Measurement Waterline Lengths (MWL) against
opponents of various lengths (January Perth conditions), from Oliver et al. (1987)
· 90 ·
ACROBAT
The detail and complexity of Sail America’s RMP2 was extended by the ACROBAT
program, developed in conjunction with Team New Zealand for the defence of the
2000 America’s Cup.
ACROBAT attempts to create an accurate simulation of a match race by modelling

the performance of the two yachts, the interactions between them, the course on
which they sail, and the wind conditions they encounter.
ACROBAT incorporated the following features:

• fixed-time-increment simulation, with integration of accelerations between
time steps;
• wind velocity and direction modelled as Markov chains, utilising the Taylor
hypothesis for movement of the wind field. This treats wind turbulence,
although random, as fixed, with eddies travelling downwind at a mean wind
speed , so an anemometer a distance d upwind of a mark, will give exactly the
same reading as an anenometer at the mark ( / ) seconds later.
• variation of wind fields between boats, correlated to their separation;
• a highly detailed tactical decision model for determining the course taken by
the yachts, including:
- simulation of course and velocity changes during tacking;
- mark rounding simulation;
- wind shadow and backwind effects between yachts.
ACROBAT tactical decision model
ACROBAT included a complex set of penalties in order to encourage each yacht to

sail an optimal route. Penalties were applied for:
• sailing in the disturbed air from a leading boat;
• sailing on the headed tack;
• lateral separation from a trailing boat;
• straying too far from the centre of the course;
• sailing past a layline on the wrong tack;
• tacking when below equilibrium speed;
• approaching the mark from the incorrect side;
• collision courses with right-of-way yachts;
• passing situations.
· 91 ·
Penalties within ACROBAT are weighted and summed to give a total penalty for
the current tack:
(5.4)
where are weights for each penalty.
The total penalty for the opposite tack is given by:
(5.5)
The tack with the smallest total penalty is chosen to be the best tack at that point in
the simulation. However, the tactical model as described by Philpott does not
appear to look ahead and, as a result, may be prone to sailing into tactical situations
less favourable than currently encountered.
Although the use of different tactical models for each boat is described in some
detail by Tierney (1998), later descriptions of ACROBAT downplay this. Philpott
(2004) states:
“In order to compare yacht designs with different design tradeoffs there must be no
bias introduced with regard to different tactical abilities of the two helmsmen.
Consequently each yacht uses an identical tactical decision model” (Philpott et al.
2004, p. 11)
5.2.3 Proposed VESPA RMP methodology
The design of the RMP for use within VESPA involves several key decisions that
determine the architecture and overall performance of the complete system. An
RMP may be a simple probabilistic model, such as that used for the Sail America
RMP1, taking a yacht performance profile derived from a VPP and combining it
with an expected wind distribution for the race, to produce a combined probability
density function (PDF). This can be integrated and compared to that of a different
yacht in order to determine the probability of winning a particular race.
The computational performance of this approach is expected to be good. However,

care must be taken to ensure that this method encapsulates sufficient detail to
accurately rank two yachts in a simulated race.
At the other end of the scale of complexity, it is possible to perform a full simulation
of one or more yachts racing around a course, involving tacking, gybing and mark
rounding manoeuvres in varying wind and sea conditions. This option involves a
great deal of computation, particularly if accelerations between time steps are taken
into account, and may be prohibitively slow when incorporated into an
optimisation procedure. This is the approach taken by the ACROBAT RMP.
· 92 ·
The primary motivation for the level of detail incorporated in ACROBAT is the
belief that it enables the program to better estimate the probability of one yacht
winning against another:
“Previous RMPs do not consider the tactical advantages that a faster yacht has over
a slower yacht… and therefore may underestimate the probability that a faster
yacht wins… In a simulation of a match-race, the tactical advantages that a
leader has can be modelled, thereby improving the estimate of the win/loss
probability when comparing yacht designs.” (Philpott et al. 2004, p.2)
However, the benefit claimed by Philpott is unlikely to have been as great as was
expected, due to an over-emphasis on factors involving interaction between boats,
such as wind shadow and backwind. While the leading boat in a match race has a
wind shadow that may affect the trailing boat and result in an increased lead, such
interactions are not as common as might be expected.
Rather than a trailing boat experiencing significant disturbed air on a windward leg,
it is the threat of disturbed air that is used by the leading boat to shepherd the
trailing boat towards a layline. Once pushed to the layline, the trailing boat has
limited tactical and strategic options and will be forced to sail in disturbed air, or to
sail extra distance to avoid it.
The losses due to interactions between yachts once both are at the layline are more
easily quantified as the relative positions of the yachts are restricted to a small
number of alternatives:
• line astern;
• opposite tacks;
• trailing boat forced above layline to obtain clear air.
Although a detailed simulation such as that used for ACROBAT may be useful for
evaluating the effects of design variations that influence manoeuvring, or for real-
time modelling of tactical situations for crew training purposes, the additional detail
included adds little to the evaluation of yacht performance. It may result in a less
accurate model that is more difficult to verify and validate. In this regard Sánchez
(2006) gives the following advice:
“Many modelers make the mistake of equating detail with accuracy. They start
with a grand vision of a highly detailed model which mirrors every aspect of the
real world system…The sheer magnitude of such programs makes verification and
validation nearly impossible. The behavior of the program is determined by dozens
to hundreds of …inputs whose correspondence to reality is tenuous at best.”
(Sánchez 2006, p.4)
· 93 ·
For a race model intended to rank the performance of two boats, the features of
ACROBAT described in Section 5.2.2 were considered unnecessary for the following
reasons:
Integration of accelerations between time steps. Prior to the Version 5 ACC rule,
a significant range of waterline lengths, displacements and sail areas could be
adopted for an ACC design. As a result, the differences in rates of acceleration
between two boats when manoeuvring or sailing in varying wind conditions may
have been significant. Under Version 5 of the ACC rule, all boats can be considered
to have the same length, displacement and sail area, and differences in their rates of
acceleration are insignificant.
This is particularly true on upwind legs due to the relatively narrow upwind speed
range speed of ACC yachts. The VMG for a typical ACC yacht varies by little more
than one knot for wind velocities from 9 to 20 knots. Changes in hull velocity as
wind velocity changes are small and the benefits of integrating accelerations between
time steps upwind are negligible. Downwind, boat speed variation across the wind
range is greater, with a 5 knot difference in downwind VMG between 9 and 20 knots
of wind. However, the rate of acceleration of the boat is also greater due to the large
amount of additional sail area carried.
Rather than using a fixed-time-increment simulation, the VESPA RMP divides the
distance between each pair of marks into 100 bands (approximately 50 metres per
band), stochastically sampling wind velocity and direction from the specified
distributions for each band. Each band is regarded as a steady-state simulation, with
yacht performance derived from the VPP. The time taken for each yacht to sail
through each fixed band of wind is calculated and added to the elapsed time for the
yacht for the leg.
Wind velocity and direction modelled as Markov chains. The VESPA RMP
randomly samples the specified wind distribution. However, as the simulation uses
discrete steps without integration of accelerations, the order of the samples is not
relevant and Markov chains for wind velocity and direction are not required.
Variation of wind fields between boats correlated to the separation between

them. ACROBAT incorporates a complex wind model that varies both spatially and
temporally. The effect is that two boats sailing close together will experience similar
wind conditions, but as the boats separate, the wind fields they encounter will differ.
This introduces unnecessary variance into the race model, blurring the outcome.
Philpott confirms this interpretation:
“In the case where the yachts see no correlation in the weather conditions it appears
that the advantage of the faster yacht has been reduced due to the random nature
of the weather observed on each yacht, which swamps the speed difference between
the yachts” (Philpott et al. 2004, p.15)
· 94 ·
As a consequence, the VESPA RMP uses identical wind fields for each boat in order
to discriminate performance differences between them without the introduction of
unquantified variance.
Highly detailed tactical decision model for determining the course taken by the
yachts. Although this is a centrepiece of ACROBAT’s design, the complexity and
uncertainty involved in a detailed tactical decision model was considered counter-
productive for the VESPA RMP. Consequently, the VESPA RMP ignores alternative
courses on each leg and the tactical considerations that may determine them. Each
leg is treated as a one-tack “drag-race”, with no need for tactical route planning and
no interactions between boats until the mark is reached.
Tacking simulation. Subsequent to the introduction of the version 5.0 ACC rule,
no appreciable difference exists between the length, displacement and sail area of
different ACC yachts. While this may not have been the case in previous years, the
difference in tacking dynamics between boats is now small. Differences that do exist
are primarily due to factors that may be difficult to quantify in a simulation, such as
the effects of keel and rudder area and section, or overall balance of the yacht.
In general, boats engaged in a match race will minimise the tacks performed and, on
average, will tend to execute a similar number of tacks during a race. As an example,
the two competitors in the seven races of the 2007 America’s Cup completed the
tacks listed in Table 5:
Table 5. Tacks Performed During Races 1-7, America's Cup 2007
Leg 1 Leg 3
Leading boat Trailing boat Leading boat Trailing boat
Race 1 15 15 17 16
Race 2 7 9 4 5
Race 3 6 5 6 8
Race 4 1 1 12 13
Race 5 2 3 8 9
Race 6 3 5 6 7
Race 7 9 9 10 9
Total 44 46 64 66
For the fourteen legs listed, three had the same number of tacks for each boat, eight
legs were won by the boat having the fewest tacks and three legs were won by the
boat executing the greater number of tacks. ETNZ, winner of two of the seven races,
performed 112 tacks, while Alinghi performed 108 tacks.
The VESPA RMP disregards both the dynamics of tacking and the number of tacks
required on the basis that they have little effect on the ranking of version 5.0 ACC
yachts. Although there is a small bias in favour of the boat that executes fewer tacks,
· 95 ·
this may be handled statistically and applied as a penalty function for the trailing
boat at the completion of each leg.
Mark rounding simulation. Each boat in a match race is required to round the
same number of windward marks in the same direction. As all ACC yachts designed
to version 5.0 of the ACC rule have effectively identical length, displacement and
sail area, these mark roundings can be considered identical, with no benefit accruing
to either boat as long as the boats are not overlapped. An example can be seen in
Figure 52, which shows almost identical paths followed by both competitors during
an actual mark rounding, involving a change of course of greater than 210 degrees.
Figure 52. Windward mark rounding, America's Cup 2007, race 5, leg 1
Examination of actual America’s Cup match races shows that overlaps at windward
marks are uncommon, due to the disadvantage suffered by the trailing boat once the
layline is reached. For example, analysis performed by the author of windward mark
deltas during the 2007 America’s Cup, showed deltas ranging between 7 and 25
seconds. At no time were the yachts overlapped at any mark roundings other than
start and finish lines during the seven races sailed.
Consequently, VESPA does not model the dynamics of either windward or leeward
mark roundings, assuming instead that marks are rounded with no advantage
accruing to either boat.
Wind shadow and backwind effects between yachts. ACROBAT explicitly models
wind shadow and backwind interactions between boats, particularly on the
windward legs. This appears to be based on the assumption that, as the leading boat
casts a wind shadow that is capable of significantly disturbing the wind encountered
by the trailing boat, it is therefore essential that this effect be modelled.
· 96 ·
This is a naive assumption based on a misinterpretation of yacht racing tactics and

strategy. In match racing, the effect on the trailing boat of sailing in disturbed air is
so great that this situation will be avoided wherever possible. Racing sailors learn
very early in their careers that clear air is paramount and will take whatever steps are
necessary to obtain it.
This can be seen clearly in traces taken from actual America’s Cup races, as shown in
Figure 53. During these windward legs, the trailing boat generally avoids disturbed
air, either by splitting tacks with the leader, as shown at position A on leg 1 of race 1,
or by establishing sufficient lateral separation on the course to be in undisturbed
airflow. This situation usually continues until both boats are at the layline, typically
100-200 metres from the windward mark.
While a trailing yacht is not always able to establish clear air, it is the exception
rather than the rule for a trailing yacht to be forced into this position for extended
periods. Situations where a leading boat may have been in a position to affect the
wind encountered by the trailing boat are those portions of the legs shown where
the tracks are close and parallel. Other than when both boats are on a layline for the
windward mark, where a trailing boat has no alternative but to accept any bad air it
receives, this situation only occurs for the races illustrated during the latter portion
of leg 2, race 2 and for a brief period in the middle of leg 3, race 3.
Remarkably, rather than the trailing boat being adversely affected by bad air from
the leading boat on these two legs, the trailing boat actually made significant gains
and was able to pass the leading boat. This can be seen in the race time data shown
in Table 6, which shows that the trailing boat passed the leading boat on both legs.
Table 6. Times for Races 2 and 3, America's Cup 2007
ALINGHI ETNZ ALINGHI ETNZ

Race 2 Gain Lead Delta Lead Gain Race 3 Gain Lead Delta Lead Gain
Leg 1 0:22 • 0:19 Leg 1 1:23 • 1:31
Leg 2 • 0:13 0:06 Leg 2 0:21 1:02 •
Leg 3 0:15 • 0:28 Leg 3 1:17 • 0:15
Leg 4 0:28 • 0:13 Leg 4 0:25 • 0:40
Despite the two boats sailing close parallel courses for portions of these legs, there is
no evidence of the trailing boat suffering from disturbed air. In fact, during the
periods of close, parallel courses, both boats established positions giving them
mutual clear air, which they were able to maintain until the last few hundred metres
of the leg. This is shown in the images on the right of Figure 53, which illustrate the
yachts at points B and C. In both positions, each yacht has clear air despite their
close proximity.
· 97 ·
Figure 53. Windward leg details, America’s Cup 2007, races 1 to 3
· 98 ·
These examples suggest that when simulating match races for the purposes of
ranking yachts, the interactions between the boats are generally not an issue until
the boats are on the same layline for the windward mark. Once in the vicinity of the
mark, the penalty for the trailing boat resulting from the wind shadow and
backwind of the leading boat can be handled statistically rather than requiring a
physical simulation. Importantly, the loss suffered by the trailing boat in this
situation is always limited to the extra distance that it must sail beyond the layline
in order to obtain clear air prior to tacking.
5.2.4 Accounting for uncertainty
The outcome of a single match-race may be influenced by a considerable degree of

random variation in the environment encountered by the yachts. Wind velocity and
direction vary both temporally as well as spatially, waves of different size and
wavelength impede the motion of the yacht, and sailors make errors in boat and sail
handling.
In order to estimate the relative performance of two yachts it is necessary to

incorporate these variations within the RMP.
Wind distributions
A wind velocity distribution for the event is defined. In its simplest form, this
consists of a mean wind velocity and standard deviation for the duration of the
event.
A series of races are defined, each having its own set of wind conditions. The wind
velocity distribution for the entire event is sampled to derive a mean wind velocity
for the race. This value is used to create a wind velocity distribution using an
individual standard deviation (SD) for that race. This SD covers only the period of
the race (2-3 hours) so is significantly smaller than the SD for the entire event.
Importantly, this SD may vary based on other factors. For example, the variance in
the wind velocity may be correlated with both the wind velocity and direction.
Conversely, two races having the same mean wind velocity may encounter different
variance in the wind conditions.
This wind velocity distribution used by VESPA for each race is clipped to upper and
lower limits. These are required due to race regulations for the Americas Cup event,
which stipulate a minimum and maximum wind velocity in which racing may take
place.
· 99 ·
Random variation in performance
VESPA uses a stochastic wind model, incorporating Monte Carlo sampling of wind
distributions for wind velocity and direction. However, once a wind strength and
direction is chosen for each step of the simulation, each race is modelled in a
deterministic manner.
When the win/loss ratio for a number of such races is used as the measure of merit
for an optimisation process, the amount by which a boat wins a race does not
contribute to its overall fitness. A race that is won by one second ranks equally as a
race that is won by ten minutes. This does not produce a robust solution, as only a
small error in the estimation of a yacht’s performance may result in a significant
change in the overall win/loss ratio.
In real-world racing conditions, the greater the boat speed advantage a yacht has
against its competitors, the better. In a tournament where random variation occurs
in the racing conditions, a large boat speed advantage will translate into a higher
win/loss ratio.
As an example of why this is the case, consider two boats, one of which is only one
second faster around the course than the other. In a strictly deterministic race model
simulation, the faster boat will win 100% of its races. Yet in the real world, the
faster boat will not win 100% of its races, nor will it win a large majority of its races.
Rather, it will win only slightly more than 50% of its races. The randomness
inherent in real-world sailing conditions will outweigh the small boat-speed
advantage in almost every race.
Conversely, if one of the yachts has a 30-second boat-speed advantage around the
course for the specific set of conditions, when the two boats race one another in
real-world weather conditions multiple times, the slower boat may still win a small
percentage of the races.
The slower boat’s win/loss ratio will not be zero because a random component exists
due to variations in wind velocity and direction experienced by each boat, variations
in the waves they encounter, as well as the inevitable errors in boat handling and
crew work. This random variation will occasionally favour the slower boat
sufficiently for it to overcome the boat speed advantage of the faster boat. However,
the faster yacht will win a large majority of the races, as its boat speed advantage will
outweigh the random variations in sailing conditions in the majority of races.
In order to provide a more robust measure of the probability that one boat will beat
another, this variability may be modelled using Monte Carlo methods. Races
simulations are repeated multiple times with identical environments, but with a
small, random time variation added to the elapsed time for each yacht for each leg
of the course.
· 100 ·
Note that if a race consisted of only one leg, it would be sufficient to treat the race
times for each yacht as distributions. In this case, the probability that one yacht
defeats another may be determined by comparing the mean and standard deviation
of their race time distributions. However, the multiple-leg format of America’s Cup
races prevents this approach being taken as each turning mark effectively applies a
small penalty to the trailing boat. The total elapsed time for each yacht is therefore
dependent on interactions between the yachts at these marks and cannot be
determined in a statistical fashion.
As the random time variation for a particular leg of a racecourse is a summation of a

large number of randomly occurring events, it is expected to have a normal
distribution. The variance of this distribution is difficult to determine through
simulation, but may be estimated from actual performance data of America’s Cup
Class yachts, such as that logged during two-boat testing.
In order to estimate the magnitude of the random noise inherent in real world
sailing conditions, data was obtained from the Alinghi design team for a series of
two boat tests, taken over a period of 15 days during May and June 2006. The
purpose of these tests is to examine the effects of small adjustments to sail trim, rig
setup and appendage configuration. Consequently, the yachts used for testing are set
up to be as similar in performance as possible.
Twenty-five test runs, which encountered wind velocities in the range of 7–15 knots,
were selected for analysis. This selection ensured that the wind velocity distribution
was similar to that expected for the America’s Cup match in June 2007 and that the
mean and standard deviation of the wind velocity were close to that used by
VESPA’s race modelling program.
The results of this testing are shown in Table 7, where the test-run date and time,
standard deviation of the true wind direction (TWD SD), VTW, VMG and boat lead at
the end of each 10 minute test run are tabulated. Results of this investigation found
that the SD of the time difference between the boats in seconds for a 10-minute test
was 4.47 seconds.
· 101 ·
Table 7. Results of Two-Boat Testing, May-June 2006
Date Start time End time TWD SD VTW VMG Lead

(yyyymmdd) (hhmmss) (hhmmss) (degrees) (knots) (knots) (seconds)
20060525 155856 160858 1.56 7.52 7.10 3.39
20060525 161234 162239 1.61 7.54 7.26 8.47
20060529 140702 141707 1.44 12.97 8.59 1.33
20060529 142654 143658 2.40 11.75 8.36 1.62
20060529 144024 145030 1.68 12.02 8.32 9.14
20060529 160658 161710 3.15 11.52 8.28 9.41
20060529 162609 163624 1.67 12.70 8.44 4.24
20060529 164010 165015 1.33 12.88 8.55 3.36
20060530 120025 121033 2.15 11.95 8.24 13.43
20060530 124147 125152 2.43 12.66 8.35 -2.41
20060530 125810 130822 2.50 13.42 8.38 -3.99
20060530 140524 141549 2.14 14.77 8.64 7.73
20060530 142410 143423 2.07 14.35 8.53 8.47
20060602 160230 161232 1.80 13.79 8.69 3.70
20060607 145500 150505 3.62 10.07 8.02 10.82
20060607 154440 155445 1.94 12.64 8.60 5.10
20060607 160657 161704 3.30 13.90 8.59 -0.82
20060607 162027 163030 0.98 14.55 8.60 2.49
20060608 113059 114447 2.59 11.30 8.31 0.28
20060608 114816 115822 1.93 11.57 8.58 8.59
20060608 142235 143242 2.10 12.73 8.42 7.43
20060608 144426 145430 1.51 13.09 8.53 -0.32
20060609 135843 140846 3.87 11.16 8.39 4.94
20060609 141210 142713 4.75 11.93 8.53 6.17
20060609 144145 145147 2.41 12.69 8.68 10.30
mean 12.22 8.36 4.92
std. deviation 1.80 0.39 4.47
The windward leg of an America’s Cup race for the 2007 event is specified as 3.0
nautical miles. Using the average VMG taken from the two boat testing results in a
windward leg that takes approximately 20 minutes to complete when sailing in
average wind conditions.
As the test period used was typically 10 minutes, this SD value needs to be scaled to
suit the longer length of legs used for actual racing. To do this, the SD can be scaled
using the equation for generalised volatility σ for time horizon , which is
expressed as:
σ √ (5.6)
· 102 ·
If the SD of the time difference between two boats for a 10-minute test period is
4.47 seconds and the SD for a 20-minute period is required, this is calculated as:
20
σ 4.47 6.32 seconds (5.7)
10
Subsequent to this analysis, VESPA’s RMP incorporated multiple repetitions of each
race with noise in the form of a random time penalty added to one boat for each leg
of the race, in order to produce a more robust estimate of a yacht’s probability of
winning races. The magnitude of this penalty is derived from a normal distribution
with a standard deviation of 6 seconds. This standard deviation, derived from the
two boat testing data, was considered conservative, yet sufficiently large to have an
effect on the outcome of the optimisation process.
Random variation in wind direction
Wind varies not only in its strength but also in its direction. Large changes in
direction occur from day to day, but smaller changes also occur on shorter
timescales, even in apparently steady wind conditions. Some wind conditions, such
as thermal sea-breezes, have a stable mean direction, but oscillate regularly by 5-10
degrees either side of the mean.
It is possible that in these conditions the strategies employed to sail the fastest
course may result in a change to the optimal design for an ACC yacht. To
understand why this may be the case it is necessary to consider a strategy that,
although widely used by sailors for many years, who referred to it as “footing to the
header”, was first formalised in 1987 by Ockam Instruments (2006) and given the
name “wallying”.
Wallying was first used by the Stars & Stripes crew competing in the 1987 America’s
Cup. Course recommendations were calculated by the yacht’s instrumentation and
navigation system, based on the polar performance curves for the yacht.
To avoid the onboard TV cameras and microphones revealing a new technique that
the Stars & Stripes crew considered a key competitive advantage, the crew referred to
a fictitious crewmember, Wally, when course adjustments were relayed to the
helmsman, Dennis Conner. Messages from the Navigator such as “Wally suggests
two tenths faster than target, Dennis” were clear to those onboard, but baffling to
those unaware of the technique being used (Teeters 2004).
If a yacht is sailing towards a mark that is directly to windward, the best

performance that can be obtained is referred to as Velocity Made Good or VMG. This
can be illustrated using a polar performance curve, as shown in Figure 54, where
best VMG is a found by a line perpendicular to the wind direction and tangent to the
polar curve.
· 103 ·
Figure 54. Polar performance curve with VMG
If a permanent wind shift occurs, and the magnitude of the shift is not sufficient to
allow the yacht to lay the mark on a single tack, the optimal course remains the
angle of best VMG.
However, when the wind direction shifts, but is expected to shift back prior to the
yacht reaching the lay line, it is possible to improve on the speed made good
towards the next mark by using a technique that involves sailing at a angle lower
than VMG.
Wallying is the practice of deviating from the optimal velocity made good to
windward (VMG) when sailing in a wind direction that is shifted from the mean
wind direction, such that the velocity made good in the direction of the next mark
(VMC) is maximized (Figure 55).
Figure 55. Polar performance curve for a shifted wind direction,

showing advantage of wallying
· 104 ·
Benefits from wallying only eventuate when the wind shifts back towards its original
direction or beyond. If the shift in wind direction is permanent, i.e. a persistent shift,
wallying will result in a longer elapsed time to reach the windward mark.
Consequently, the successful use of wallying is dependent on the probability that
the wind, once shifted in direction, will shift back prior to the yacht reaching the
layline. For this to occur the wind should be oscillating in its direction, and there
should be sufficient time for one full oscillation before the layline is reached.
Wallying is best used when the expectation of a mean reverting oscillation is high,
for example, near the start of a windward leg, when the wind is oscillating
predictably, with a period shorter than the length of the leg. Conversely, wallying is
unlikely to be used when the time to the layline is short, the wind is shifting
persistently in one direction, or when the oscillation period is long.
Wallying was included in the VESPA race model in order to investigate whether
wallying in winds of variable direction could have a measurable effect on the design
of the yacht. Because wallying requires the helmsman to sail at a slightly lower
heading and at a higher boat speed upwind, it is conceivable that a boat designed
for a regularly oscillating wind might have slightly different optimal design
parameters compared to a boat designed for the same wind velocity but without
directional variation. For example, a boat designed for winds that oscillated
significantly in direction would on average sail lower and faster, so righting moment
may need to be greater, meaning wider BWL and/or higher CM, as well as higher CP
and a LCB that is further aft.
Trailing boat penalties
The purpose of trailing boat penalties is to quantify the benefit of being the first
boat to reach a windward mark in a match race. This benefit is important as it
determines to what extent a boat should have its performance biased to upwind
work relative to downwind work.
These considerations resulted in the inclusion of two different forms of trailing boat
penalties within the VESPA race model. The first is a simple port tack penalty – if
overlapped, the port tack boat is forced to cross behind the starboard tack boat.
The second component is a penalty for disturbed air. If a boat is behind at the first
mark, it is penalized by the addition of a few seconds to its elapsed time for the leg.
This penalty is at a maximum of 6 seconds when the boats are almost overlapped,
tapering linearly to zero for a lead of 20 seconds.
Initially it was considered that a separate trailing boat penalty for leeward mark
roundings would be necessary in addition to the windward mark trailing boat
penalty. However, the course for the 2007 America’s Cup was modified from that
· 105 ·
used for the 2003 America’s Cup by the provision of a leeward mark gate, made up
of two marks approximately 150 metres apart.
The effect of this was to remove any penalty for the boat trailing at the end of the
downwind leg. Instead of being forced to follow the leading boat around the mark,
making it easy for the leader to gain tactical control, the trailing boat could opt for a
different mark than that rounded by the leading boat. As a result of this change to
the design of the course, it was concluded that leeward mark roundings did not
require a trailing boat penalty to be applied by the VESPA RMP.
5.2.5 Summary
Rather than attempting to create a perfect simulation of the physics of a match race,
the VESPA RMP focuses on modelling the simplest possible race that will correctly
rank two boats. Factors that help to clarify the primary performance differences
between yachts have been given priority, while second order effects, such as the
impact of hull shape changes on manoeuvring, are ignored.
Some of this simplification is possible because of recent changes to the ACC rule. In
previous versions of the ACC rule a range of waterline lengths, displacements and
sail areas were legal. The version 5.0 amendments to the ACC rule reduced these
ranges to the point that all boats can be considered to have the same length,
displacement and sail area. Consequently, there now are negligible differences
between yachts accelerating and decelerating in gusts and lulls, or during tacking or
mark rounding manoeuvres.
As short execution times are considered essential, the VESPA RMP dispenses with
direct physical modelling of many of the complexities of yacht racing where it is
considered that no net benefit for either boat occurs, or when the effect can be
approximated in a probabilistic manner. The VESPA RMP has been designed to have
exceptionally low execution times, as the optimisation approach adopted may
require that millions, rather than hundreds, of race simulations be performed.
· 106 ·
Chapter 6 · Optimisation Algorithm
6  Optimisation Algorithm
Chapter 4 described the use of a neural network metamodel, which approximates
hydrodynamic data derived from the SPLASH potential flow code for variations to a
parent ACC hull design.
In order to find the best design within the parameter space, it is necessary to explore
this space using some form of optimisation algorithm. This chapter examines the
various optimisation methods available and describes the factors determining the
selection of a particular optimisation methodology for VESPA.
6.1 Optimisation methods

Many different approaches to non-linear parameter optimisation exist. These can be
broadly divided into three categories:
• Gradient based approaches. These methods calculate first, and in some cases,
second derivatives of the function being optimised, in order to approximate
the function to determine the next step in the optimisation process. Examples
include the Newton-Raphson, quasi-Newton, Levenberg-Marquardt and
conjugate gradient methods.
• Pattern search methods do not rely on derivative information, relying instead
on comparison between multiple evaluations of the function performed in a
· 107 ·
fixed pattern. The best known of these methods are the Nelder-Mead downhill
simplex and the Hooke-Jeeves pattern search.
• Stochastic optimisation methods are those that introduce some element of
randomness in their search for an optimal solution, rather than proceeding in
a totally deterministic way. These methods include simulated annealing,
particle swarm optimisation and evolutionary algorithms.
Given this degree of choice, the selection of the most suitable optimisation method
is typically determined by the availability and reliability of the derivative
information; the presence of noise in the function; and the potential for the
function to be multimodal.
In the case of ACC yacht design optimisation, several other factors restrict the
choice of optimisation method. The use of a competitive fitness function, i.e. one
that is based on a comparison of the performance of one yacht against a variety of
others, requires an entire population of boats to be modelled and improved. When
the fitness of a design is based on its performance relative to its competitors, the
fitness landscape is dynamic, changing with each step in the optimisation. The
optimisation method adopted should be capable of handling such a dynamic fitness
landscape.
In the case of VESPA, the optimisation algorithm needs to meet the following
requirements:
• capable of handling non-linear and multi-modal solutions;
• able to evaluate competitive fitness functions;
• capable of finding solutions to problems having dynamic fitness landscapes.
6.2 Gradient and pattern search methods

Although the need for a population of solutions and the presence of a dynamic
fitness landscape bring with them some challenges, they do constrain the choice of
optimisation method. The dynamic nature of the fitness landscape makes it unlikely
that a gradient based method (e.g. Newton-Raphson or Quasi-Newton) or search
pattern based method (e.g. Nelder-Mead simplex or Hooke-Jeeves direct search)
would behave in a stable fashion. The fact that a population of boats needs to be
modelled and evaluated makes stochastic, population-based approaches, such as
Genetic Algorithms and Particle Swarm methods a natural choice.
Consequently, traditional gradient based or search pattern based methods can be

eliminated from the search for a suitable optimisation method for this research. This
is not to say that these methods are not applicable if the structure of the objective
function differed. For example, if a Volvo 70 yacht design intended for an around
· 108 ·
the world yacht race was being optimised, the measure of merit would be concerned
with elapsed time on the course rather than performance against a single opponent.
In this case, a gradient-based optimisation method may be faster and give more
accurate results.
6.3 Stochastic methods

There are three widely used approaches to stochastic optimisation, as illustrated by
Figure 56. Within the category of Evolutionary Algorithms the taxonomy branches
further to include three subgroups.
Stochastic
Optimisation
Methods
Simulated Evolutionary Particle Swarm

Annealing Algorithms Optimisation
Genetic Evolution Genetic

Algorithms Strategies Programming
Figure 56. A taxonomy of stochastic optimisation methods
6.3.1 Simulated Annealing
Simulated Annealing (SA) is a global optimisation method analogous to annealing

in metallurgy, the controlled cooling of a material to increase the size of its crystals.
At higher temperatures, atoms are free to move from their initial energy states and
randomly move through states of higher energy; slow cooling increases the
probability of finding configurations with lower internal energy than the initial one.
The SA algorithm proceeds by replacing the current solution with a nearby one,
chosen at random with a probability that depends on a global temperature value T.
At the commencement of the optimisation process, T is large, allowing greater
movements in the parameter space and permitting solutions that may be worse than
the current one. As the optimisation proceeds, T is reduced resulting in smaller
steps and a smaller allowance for worse solutions. The allowance for solutions that
may be worse than the current one helps to prevent the method becoming stuck in
local optima, and it is this feature that distinguishes SA from a simple search
method.
· 109 ·
SA was found independently by Kirkpatrick et al. (1983), and by Černý (1985) and
is an adaptation of the Metropolis-Hastings algorithm, described by Metropolis et al.
(1953). Chen (1996) utilised a simulated annealing algorithm to optimise the
principal dimensions of a ship, noting that the method was able to avoid a local
extreme point and find the global optimal point, regardless of the choice of initial
point. Morishita and Akagi (2005) found significant reductions in the calm water
resistance of both monohull and catamaran fast-ferry hulls, using a simulated
annealing method.
Although SA is by definition a stochastic rather than deterministic optimisation

method, it optimises a single design rather than being population based, making it
unsuitable for the population-based measure of merit required for VESPA.
6.3.2 Particle Swarm Optimisation
The Particle Swarm Optimisation (PSO) method was created by Eberhart and
Kennedy (1995). A stochastic, population-based method, PSO was inspired by the
behaviour of a flock of birds flocking or a school of fish. Potential solutions to the
problem under investigation, referred to as particles, are initially placed randomly
throughout a multi-dimensional parameter space. Each particle has three pieces of
knowledge: the best solution that it has found along its trajectory through the
parameter space; the best solution encountered in the particle’s neighbourhood; and
the global best solution found by any other particle.
These particles are set in motion and the direction and velocity of the motion of the
particles is influenced by the particles which have achieved the best solutions to the
problem being evaluated according to the following rules:
( ) (6.1)
(6.2)
Where is velocity, is position, is the inertia weight, is the dimension, and

are positive constants, and are random functions, is personal best and
is global best.
The algorithm proceeds through a series of time steps, updating the location and
evaluating solutions for each particle. The size of the swarm and its initial
distribution determines how effectively the global space is explored, while the
communication between members of the swarm will result in eventual convergence
on the best solution found. PSO methods are relatively simple to implement as there
are no mutation and recombination operators involved as there are with
evolutionary algorithms.
· 110 ·
Despite apparent advantages, PSO methods have not shown themselves superior to
well implemented genetic algorithms for single-objective parameter optimisation
problems. Tan (2003) found that GA outperformed PSO methods on single
objective aerodynamic shape optimisation problems, although for multi-objective
problems the PSO method was better able to locate the Pareto front. Pinto et al.
(2005; 2007) successfully used PSO methods for multi-objective optimisation of
merchant ships and found that PSO found a greater number of Pareto optimal
solutions.
6.3.3 Evolutionary Algorithms
Evolutionary algorithms are search methods inspired by the process of evolution

that occurs in biological species. Originating in research by Holland into cellular
automata, (Holland 1975), evolutionary algorithms quickly differentiated into three
disciplines, Genetic Algorithms (GA), Evolution Strategies (ES) and Genetic
Programming (GP).
All three methods share the key features of an evolutionary algorithm –

• Population based. The methods use a population of candidate designs, each
represented as a vector of real valued or binary numbers, or as a tree structure.
• Ranking. A procedure that evaluates and ranks individuals from the
population using an objective function.
• Selection. A selection procedure that chooses individuals, based on their
ranking or fitness, to have their genes represented in the following generation.
• Recombination. The recombination of characteristics from selected
individuals to create new individuals for the following generation.
• Mutation. Small variations made to the genome on a random basis in order to
introduce diversity into the population.
Although there are many similarities between the methods, the differences between
them are also significant. Genetic programming, based on the evolutionary
programming of Fogel et al. (1966) and refined by Cramer (1985) and Koza
(1992), is primarily concerned with the evolution of computer programs.
Rather than using a simple vector of numbers, genomes for genetic programming
tend to be based on tree structures where each node is a small segment of code.
Although a very powerful approach for particular problems, GP is not applicable to
the problem under investigation and can be eliminated as a candidate for an
optimisation method.
· 111 ·
There are two major philosophical differences between the fields of GA and ES; the
format used to encode parameters in the genome and the emphasis placed on the
recombination and mutation operators.
Encodings
Evolution Strategies, (Rechenberg 1973; Schwefel 1974), were designed from the
outset to use a floating-point representation. Conversely, genetic algorithms initially
used a binary representation, as the building block hypothesis and its associated
schema theorem, which had been formulated as theoretical underpinnings for
genetic algorithms by Holland (1975), implied an incompatibility with floating-
point representations.
Although the arguments by Holland and Goldberg (1989) in favour of binary

representations dominated genetic algorithm research for many years, more recent
work has cast doubt on the theoretical advantages of binary representations. Janikow
and Michalewicz (1991) and Wright (1991) demonstrated that real coded genetic
algorithms outperformed binary in a majority of test problems. Real coded GAs
have become widely used, particularly for the solution of parameter optimisation
problems where a binary representation would be difficult due to concerns related
to magnitude or precision.
Mutation and recombination
Although canonical Evolution Strategies emphasise the use of mutation as the

primary search operator, recently developed ES have introduced more powerful
recombination operators. These developments in ES, when combined with the
increasing use of floating-point representations for GA, have served to blur the
traditional distinctions between the two approaches.
The different emphasis placed on mutation and recombination remains the primary
difference between real-valued GA and ES. ES typically use high rates of mutation,
with highly specialised mutation operators, together with low levels of crossover. In
contrast, a typical GA will have higher rates of recombination coupled with low
mutation rates.
Optimisation algorithm selection
In a comparison of the performance of GA, ES, GP and SA methods for two

deceptive test problems, Keane (1996) demonstrated that basic implementations of
GP, ES and SA were all outperformed by a GA.
Previous research into the optimisation of marine vessels has indicated that the
solution space is often non-linear and multimodal. Several researchers, including
· 112 ·
Sahoo (1997), Yasukawa (2000), Hirata (2004) and Pinto (2004) found that
optimisation algorithms based on gradient information or local pattern searches
achieved inferior results to global optimisation methods, for problems related to the
hydrodynamics of marine vessels.
Based on the factors described, a floating-point based genetic algorithm was chosen
as the optimisation method for use by VESPA. Some mutation and crossover
methods from the field of ES have also been adopted. The selection of the various
parameters, operators and settings used by the VESPA GA are described in the
following section.
6.4 Proposed GA methodology

The steps involved in a genetic algorithm are as follows:
1: Initialise the population

2: Evaluate initial population
3: REPEAT
4: Perform competitive selection
5: Apply genetic operators to generate new solutions
6: Evaluate solutions in the population
7: UNTIL some convergence criteria is satisfied
Each iteration of a GA involves a competitive selection that penalises poor solutions.

Solutions with high fitness are recombined with others to produce members of the
next generation, each member inheriting properties from its parents. Solutions are
also mutated by making small, random changes to one or more free parameters.
The effect of recombination and mutation is to generate new solutions, which are
biased towards regions of the solution space from which good solutions have already
been seen.
6.4.1 GA parameters
A GA based search can be viewed as a trade off between the maintenance of

sufficient diversity in the population to permit the coverage of the entire solution
space (exploration), and the need to converge on the optimal solution within an
acceptable time (exploitation), (Eiben and Schippers 1998).
There are several variables or algorithms that can be adjusted to improve the
performance of the GA, or to manipulate the degree of exploration or exploitation
that the GA exhibits, (Goldberg 1989).
· 113 ·
These include:
• recombination
• mutation
• selection
• population size
• fitness sharing
• elitism
6.4.2 Recombination
The theory of recombination in biological organisms was first detailed by Morgan

(1916), who described both single point and dual point crossover in simple
chromosomes, as shown in Figure 57.
Figure 57. Single-point and dual-point crossover, from Morgan (1916)
Recombination within a binary genome of a GA is a similar procedure. A crossover

locus is selected at random and the bits on one side of this locus are exchanged
between the two parent genomes (Figure 58). This crossover is directly analogous to
the recombination that occurs in sexual reproduction within single chromosome
(haploid) organisms.
Figure 58. Binary recombination and mutation
· 114 ·
Crossover for floating-point representations presents a more difficult problem.

Although crossover can be performed by exchanging portions of two floating-point
vectors, this does not change individual values for the exchanged vector elements. A
binary representation, on the other hand, may perform crossover at any point in the
binary string. If the genome is made up of a vector of 8-bit integers, crossover can
occur at the integer boundaries, or within the integer itself, resulting in the changes
to the values of numbers represented as bit strings. For vectors of floating-point
numbers, crossover can only occur at the level of the numbers themselves, not
within their binary representations.
Many schemes have been proposed for recombination operators for real-valued GAs.
These include operators that exchange values between chromosomes:
• Single-point crossover: a crossover location 1, 2, … , 1 is randomly
chosen. Two new offspring chromosomes and are created from parent
chromosomes and :
( , ,…, , ,…, )
( , ,…, , ,…, )
(6.3)
• Dual-point crossover: two crossover locations , 1, 2, … , 1 with
are randomly chosen. Two new chromosomes are created:
, ,…, , ,…, , ,…,
, ,…, , ,…, , ,…, (6.4)

• Discrete crossover: is a randomly selected value from the set ,
A recombination method for a real-valued GA may also perform a weighted blend

between two or more chromosomes:
• Line crossover: an offspring ( ,…, ,…, ) is generated, where is a
randomly chosen value on the interval , .
• Extended line crossover, (Muhlenbein and Schlierkamp-Voosen 1993):
( ) (6.5)
where is a randomly selected value on the interval 0.25,1.25

• Extended intermediate crossover:
( ) (6.6)
where is a randomly selected value on the interval 0.25,1.25
· 115 ·
• Directional crossover, (Yamamoto and Inoue 1995b):
· ( )·( ) · ( )·( ) (6.7)
where and are random values on the interval 0.0,1.0 and is the value of
the fitness function for the corresponding vector of variables .
Yamamoto and Inoue define as belonging to generation , while and
are its parents from generation 1. However, and may also be derived
from generation .
Directional crossover creates a bias for the offspring in the direction of the
parents with the greatest fitness. This bias, combined with the use of three
parent points forming a simplex in the search space, introduces elements of
pattern search methods, with evident similarities to the Nelder-Mead Simplex
method. While this may speed convergence, it may also sacrifice some of the
exploration capabilities of the GA.
The recombination methods selected for use by the VESPA GA are dual-point
crossover and directional crossover. These are applied in the proportion 80:20 to
avoid premature convergence resulting from high levels of directional crossover.
6.4.3 Mutation
Mutation is straightforward when using binary representation, achieved by random

flipping of individual bits within the bit string that forms the genome.
Recommendations for the probability of mutation for each individual bit have been
determined empirically by several researchers, including de Jong (1975) and
Schaffer et al. (1989), and are typically in the range 0.005 – 0.01.
The implementation of mutation operators for real-coded GAs is not as simple as

for binary representations. Mutation of a floating-point vector genome may be
performed by randomly perturbing individual elements within the vector. As with
the binary representation, vector elements are selected using a mutation probability.
This value is typically greater for the floating-point number than would be used to
mutate the individual bits that would be used to represent that number in a binary
genome.
Once a vector element has been selected, its value has a small random perturbation
added to it. This perturbation is generated from a normal distribution using a
standard deviation specific to the vector element.
Levels of mutation that are too high can seriously disrupt the evolution process,
while low mutation levels can slow convergence considerably, particularly in
algorithms such as ES that do not emphasise recombination operators. Mutation for
· 116 ·
real coded parameters has been extensively researched by the ES community with
schemes for adaptive mutation and covariance matrix based mutation widely used.
The VESPA GA uses random mutation with probability of 0.1 per design variable.
Perturbations are sampled randomly from a normal distribution having a standard
deviation equal to 0.05 times the parameter range (i.e. upper limit – lower limit) for
that design variable.
6.4.4 Selection
Much research has been done into different selection methods with two of the most
widely used methods being roulette wheel selection and tournament selection.
Roulette wheel selection assigns a likelihood of representation in the following

generation proportional to the magnitude of an individual’s fitness score. The
difficulty inherent in the use of roulette wheel selection is that the likelihood of
representation in the following generation is dependent on the scaling factor used to
scale the results of the objective function to create fitness scores. Changing this
scaling factor can dramatically change the results of the selection process, and it is
difficult to determine the correct scaling factor ahead of time.
Tournament selection works by choosing two individuals from the population and
comparing their fitness scores. The individual with the highest fitness score is
selected to be a parent for an individual in the subsequent generations.
As tournament selection is a simple comparison of fitness values, it is independent

of scaling factors. Tournament selection is both easier to implement and more
robust in use than roulette wheel selection.
Consequently, tournament selection has been adopted for use by the VESPA GA due
to its inherent simplicity and robustness. The absence of scaling issues is seen as a
distinct advantage in an environment where the range of objective values may vary
dramatically due to the dynamic fitness function used by VESPA.
6.4.5 Population size
For a GA, the smaller the population, the faster each generation will run. On the
other hand, it is vital that the population size is large enough to maintain sufficient
diversity for evolution to progress.
Despite the widespread use of Genetic Algorithms there seems to be little agreement
on the ideal population size. A review of the literature reveals population sizes
between 10 and 300 in use, with few researchers providing justification for their
choice of population size. Grefenstette et al. (1986) studied optimal values for
· 117 ·
control parameters, reporting that a population size of 30 gave the best results.
Similarly, Schaffer et al. (1989) obtained the best results using a population size of
20 to 30 individuals.
In previous work by the author (Mason et al. 2005) a sensitivity analysis was
performed for population sizes ranging between 10 and 150, for a multimodal, non-
linear problem that involved finding the minimum wave resistance of a catamaran.
Negligible differences in the rate of convergence or the fitness level of the best
solution were observed for populations greater than 20 individuals.
6.4.6 Fitness sharing
Fitness sharing encourages diversity in the population by scaling the fitness of

individuals based on how similar they are to all other individuals in the population.
Individuals that are very similar to others in the population are given a slight penalty,
while individuals that show novel features are rewarded by increasing their fitness
relative to the population. This is intended to ensure that the population maintains
sufficient diversity to avoid premature convergence on false optima.
Fitness sharing was shown by Goldberg and Richardson (1987) to be effective in

maintaining population diversity, which helps to prevent premature convergence on
local rather than global optima. While this may be of use during the exploration
phase, this diversity may inhibit convergence to an optimal solution during the
exploitation phase.
Although the roulette wheel selection method appears to benefit from fitness
sharing, Oei et al. (1991) found that fitness sharing can display chaotic interactions
when combined with tournament selection.
While fitness sharing has some attractive properties, the potential for chaotic
interactions with the dynamic fitness function and the tournament selection
method used by VESPA outweigh the benefits. Fitness sharing has not been adopted
for use with the VESPA GA at this time.
6.4.7 Elitism
Elitism ensures that the next generation’s best individual will be at least as good as
any from the previous generation by automatically including the best individual
from the previous population.
De Jong (1975) found that elitism improved the performance of a genetic algorithm
for unimodal functions, while performance for multimodal functions was degraded.
This suggests that while elitism improves local search and may be appropriate
· 118 ·
during the later exploitation phase of an optimisation, it may also inhibit global
exploration.
Rudolph (1994) showed that while genetic algorithms without elitism are not
guaranteed to converge to a global optimum, a GA with suitable mutation operators
and elitism is guaranteed to reach a global optimum if given sufficient time.
VESPA applies simple elitism to all generations of the genetic algorithm. In addition,
VESPA uses more complex elitism methods, necessitated by the competitive fitness
function adopted.
Competitive fitness
An objective function that is based on a comparison of an individual in competition

with a variety of other individuals from the population, is termed a competitive
fitness function. Competitive fitness indicates that the problem has a dynamic
fitness landscape that may change from one generation of an evolutionary algorithm
to the next.
Competitive fitness commonly takes one of two forms:

• one-population competitive fitness;
• N-population competitive fitness, also referred to as competitive co-evolution.
This form encompasses competition between multiple species, including
parasite-host and predator-prey co-evolution.
Competitive fitness has advantages and disadvantages. On the positive side,

competitive fitness can act as a self-adapting mechanism to increase problem
difficulty, as individuals in the population adapt themselves to the co-evolutionary
environment. A co-evolutionary arms race can also act to speed up the evolution
process significantly.
On the other hand, competitive fitness may suffer from problems that serve to
inhibit convergence of the genetic algorithm. These problems include:
• disengagement occurs when the gradient of the fitness values for the
population flatten, leading to stalling or drifting of the evolutionary process.
This may be accompanied by a loss of diversity in the population caused by
premature convergence on a solution, which may be mistaken for a global
optimum.
• cycling is a repeated pattern of visitation to the same areas of the solution
space. Cycling is most likely to occur if the solution contains intransitivities i.e.
B defeats A, C defeats B, but then A defeats C. However, de Jong (2004)
demonstrated that cycling can also occur for purely transitive relationships.
· 119 ·
These pathologies may be inhibited by the use of strategies that encourage diversity
in the population such as fitness sharing. One approach that has been favoured is
the use of a “hall-of-fame” (Rosin and Belew 1996), which saves good individuals
from previous generations.
The performance of an evolutionary algorithm based on competitive fitness is also

dependent on the topology of the competition adopted. Panait and Luke (2002)
identified several different structures for one-population competitive fitness:
• round robin
• random pairing
• K-random pairing
• single elimination tournament
The choice of competition topology determines the type of fitness assessment that is
used. Single elimination tournaments result in a duel methodology: A is better than B
if and only if A usually beats B in a match. Round robin tournaments result in the
Renaissance man methodology: A is better than B if A beats more competitors than B,
even if B usually beats A in a match.
Some sporting contests use a combination of the two methodologies. For the
America’s Cup, challengers compete in a round robin competition to qualify for a
single elimination tournament for a small number of boats. The winner of this
tournament then competes in an elimination match with the defender of the
America’s Cup.
The implication for ACC design optimisation is that, as the tournament structure is
asymmetric, in that the challenger and defender experience different paths to the
final elimination match, the optimal design for a challenger may not be the same as
for the defender.
Exemplars and the “hall-of-fame”
In order to prevent the occurrence of disengagement and cycling, two related

measures, based on the “hall-of-fame” approach of Rosin and Belew, were adopted
for use by the VESPA GA.
A simple hall-of-fame approach involves the inclusion in the population of the best
individual from each of the previous generations, effectively the extension of elitism
to encompass all generations. This approach is not guaranteed to be successful, as
the hall-of-fame members may not be sufficiently numerous or diverse to prevent
cycling or disengagement.
· 120 ·
An alternative approach is the inclusion in the population of an additional elite

individual that has been selected based on an objective measure other than the
competitive fitness function.
For VESPA, one candidate for this objective measure is the average elapsed course
time for a yacht in the absence of competitors. While the boat with the lowest
average elapsed time is not likely to be the boat that wins the most races, due to the
bias introduced by trailing boat penalties, this boat is always likely to be a strong
competitor, and its presence should assist to prevent cycling of the solution.
A second strategy was also adopted in an attempt to avoid cycling of the solution.
This strategy involved the creation of individuals with specific design variables,
termed exemplars, which are included in the round-robin tournament used by the
GA. For a design to be considered optimal, it must be capable of outperforming
these exemplars.
The use of exemplars is effectively a manual “hall-of-fame” method. By specifying

the exemplars directly, the method has the advantage that the design parameters of
specific competitors can be added to the tournament, allowing comparison and
evolution of proposed designs against known opponents.
· 121 ·
Chapter 7 · Implementation
7  Implementation
In order to create VESPA, a workable optimisation system for the design of ACC
yachts, it was necessary to create and integrate several key functions:
• a powerful parametric transformation function to allow VESPA to vary hull
shapes based on a small number of key parameters, while still producing hull
shapes that are both acceptable to the design team and legal under the ACC
Rule, (ACC 2003);
• sampling of the design space using a modern Design of Experiments (DOE)
method. The parametric transformation function is used to create hulls,
derived from a specified parent model, matching these sampled parameter
values;
• analysis of lift, drag and hydrodynamic heeling moment for the hulls in the
systematic series, using the SPLASH potential flow code;
• the use of neural networks to fit the SPLASH data for all hulls in the systematic
series to produce a global regression model, or metamodel, for lift, drag and
other parameters;
• customisation of a proven Velocity Prediction Program (VPP) to read the
neural network based metamodel and to allow it to be called as a Dynamically
Linked Library (DLL) from another application;
· 122 ·
• a stochastic Race Modelling Program (RMP) that creates a tournament of races

for the population of boats. This RMP uses performance data derived from the
VPP to race each boat against every other multiple times, in wind conditions
sampled from a specified distribution of wind velocity;
• a Genetic Algorithm (GA) based optimiser that takes the parent ACC hull
design, which is used to generate the lift and drag models, and evolves a
population of variations using the parametric transformation function.
7.1 Integration of components

The components of VESPA are integrated in the manner illustrated in Figure 59.
There are two main parts to the system, the creation of the metamodel, which
occurs in advance of the optimisation process being performed, and the
optimisation loop itself.
Figure 59. VESPA logic flow
· 123 ·
7.2 Parametric transformation function

Analysis of the restrictions imposed by the version 5.0 ACC rule, combined with
discussions with the key members from the Alinghi design team, determined that
the design variables with the most scope for optimisation by VESPA were topside
flare, BWL, Tc, CM, LCB and CP.
The parametric transformation function was developed in two stages:

• Stage one allowed generic parametric transformation of marine hull shapes. It
performed linear scaling of LWL, BWL, Tc and ∇, as well as non-linear variations
of LCB and CP. As both LCB and CP are coefficients (LCB is specified as a fraction
of waterline length) these are invariant under the subsequent linear
transformations required to achieve the desired LWL, BWL, Tc and ∇ values.
Note that although any three of the four linear scaling parameters (LWL, BWL,
Tc and ∇) may be specified, it is not possible to simultaneously specify all four,
as this leads to an over-constrained problem. If less than three of the
parameters are specified, these are varied such that the ratio between the free
variables is kept constant. For example, if LWL and ∇ are set to specific values
while BWL and Tc are left as free variables, BWL and Tc will be scaled accordingly
while keeping WL constant.
• Stage two implemented transformations specific to the additional design

variables required by VESPA for ACC yachts. These variables were topside flare
and CM, as well as three constraints specific to the ACC rule, LBG, FG and AG.
Hull shape convexity, another requirement of the ACC rule, was also required
as a constraint.
Although ACC yachts are restricted in length, this is measured in the LBG
plane, placed 200 mm above the datum waterline, rather than along the
datum waterline. This has a significant effect on the design of the parametric
transformation function, as ∇ is no longer a linear scaling of LWL, BWL and Tc.
The effect of measuring length in the LBG plane is that length ceases to be
independent of changes to LCB or CP. For example, while keeping the length
measure constant, a change to CP that resulted in a small change to ∇ could
now not be compensated for by making a single, proportional change to BWL
or Tc; instead, an iterative search must be conducted to find the correct value.
The non-linear form variations, such as LCB and CP, require a search procedure to
find the hull shape that satisfies the design parameters specified. This was initially
handled using a Nelder-Mead pattern search. However, the additional demands
imposed by the non-linear transformations required to satisfy flare, LBG, CM, FG
and AG values made it necessary to implement a Newton-Raphson solver in its
place.
· 124 ·
The resulting parametric transformation, when combined with a parent model of

sufficient quality, provides the ability to create a wide range of realistic, class legal
hull shapes based on a very small number of meaningful parameters. In the current
implementation, the variables chosen were BWL, CP, CM, flare and LCB, with FG, AG,
LBG constrained to within 1 mm of their maximum values, and displacement
constrained to within 1 kg of 24,000 kg.
7.2.1 Hierarchical Free-Form Deformation
To perform a parametric transformation using a hierarchical FFD, a hull design

represented by a single NURBS surface is required. Such a surface has a control-
point network having 1 rows and 1 columns. The bounding box of the
projection of each control point column into the ( , ) plane is mapped into the
unit square, as shown in Figure 60:
Figure 60. Control points mapped into the unit square
Each control point in the column has its parametric co-ordinates ( , ) in the unit
square calculated from its ( , ) co-ordinates using:
, ( )
,
( ) ( )
, ( )
, (7.1)
( ) ( )
For each control-point column, a planar bi-linear NURBS surface is created whose
corner control point ( , ) co-ordinates match those of the corners of the bounding
box for that column, and whose co-ordinates are equal to the mean value of the
co-ordinates of the control points in that column. Such planar NURBS surfaces are
· 125 ·
Figure 61. Bounding boxes displayed for each control point column in C
This array of planar NURBS surfaces is enclosed by a tri-variate NURBS volume,

as shown in Figure 62. This NURBS volume is defined by a lattice of control points
, , , having 1 control points in the direction, 1 control points in the
direction, and 1 control points in the direction. In this case, 2, 1,
and 1 , with degree 2, 1, 1.
Figure 62. FFD planes enclosed by FFD lattice P
The ( , , ) parametric values for the corner points of each planar NURBS surface
are calculated using:
(7.2)
Changes to the ( , , , ) values of the FFD control points result in an amended

FFD lattice co-ordinates , , and amended lattice weights , , . These changes
result in deformation of the embedded planar NURBS surfaces, as shown in Figure
63, which can be calculated using:
∑ ∑ ∑ , ( ) , ( ) , ( ) , , , ,
(7.3)
∑ ∑ ∑ , ( ) , ( ) , ( ) , ,
· 126 ·
Figure 63. Deformed FFD lattice P’ with updated weights , resulting in

deformations to the embedded planar NURBS surfaces Q’
The updated control point values for each modified planar NURBS surfaces ,
together with modifications to the planar NURBS weight values , are used to
create an updated hull-surface control point network .
∑ ∑ , ( ) , ( ) , ,
( , )
∑ ∑ , ( ) , ( ) ,
( ) ( ) (7.4)
The result is a hull surface that has been smoothly deformed to vary BWL, CP, CM,
flare, LCB, FG and AG in a controlled manner. An example of such a transformation
is shown in Figure 64.
Figure 64. Deformed hull surface defined by updated control point network C'
This transformation was achieved by changing only ten values:

• the co-ordinates of five control points in the FFD lattice;
• two FFD weight values, one applied simultaneously to the four forward FFD
lattice points and the other applied simultaneously to the four aft FFD lattice
points;
• three weight values applied simultaneously to each embedded planar NURBS
surface. These local weight changes are used to change the sectional shape of
the hull.
· 127 ·
7.3 Systematic series creation

To create a suitable metamodel on which an optimisation process can be based, it is
necessary to sample the design space to obtain a set of values which span the ranges
of the design variables required. These sampled values define a systematic series of
hull forms such as those shown in Figure 65.
Figure 65. A simple systematic series of hulls, varying in CP, BWL, CM , T and flare
Sampling based on either regular grids or random samples display significant

problems when used for the fitting of metamodels. It is therefore necessary to use a
form of sampling that has superior performance to both a regular grid and a truly
random sample. These methods are known as quasi-random samples, or in some
cases, as low discrepancy sequences, and display distributions that are highly
uniform in all dimensions.
In order to create the hull forms required from these sampled parameters, the
parametric transformation described in Section 7.2 is used to transform a parent
hull to match the required values.
· 128 ·
An Excel, spreadsheet shown in Figure 66, was created to generate families of hull
shapes by, deforming a parent hull form to match the parameters required.
Figure 66. Hull-series generation spreadsheet
The spreadsheet consists of four parts:

• Section A allows upper and lower limits for each design variable to be specified.
The independent variables used for the creation of the systematic series are BWL,
CM, CP, LCB, topside flare.
• Section B lists the Uniform Design matrix for 25 runs (i.e. designs) and 5
factors (i.e. design variables). This matrix was derived from uniform design
tables provided online (Fang 2004), and forms a five dimensional Latin
hypercube, with points uniformly distributed in each dimension.
• Button C runs an Excel macro to populate the systematic series shown in
Section D and creates hull geometry using the parametric transformation
function. This macro calls the VESPA parametric transformation function via a
Microsoft COM interface.
· 129 ·
• Section D lists the design variable values created by combining the values in
sections A and B using:
, 1
, ( ) ( ( ) ( , )) (7.5)
( ) 1
where is an array of design variable values, is an array of values derived

from the Uniform Design, is an index to the variable type, is an index to
the design number, while and are the upper and lower limits for
each design variable type.
Once the parametric transformation has been performed, the resulting hull
definitions for the systematic series are exported as IGES NURBS surface files. These
are subsequently used as input to SPLASH, a potential-flow CFD program.
In addition to the initial series of 25 hull shapes, two additional groups of hulls were
created. The first group of 7 hulls was created using design variables chosen at
random within the defined design space at the same time as the initial series, and
analysed using SPLASH. The SPLASH results for these hull shapes were not used for
neural network training, being reserved solely for use as an independent test set for
validation of the prediction accuracy of the neural network metamodel.
The second group of designs, also 7 hulls, was defined once a significant amount of
optimisation had been performed using the initial neural-network metamodels. This
group of designs was created using a test matrix based on the Uniform Design, but
using ranges for each design variable that were restricted to approximately 20% of
the initial range, centred on the values found to be optimal by VESPA.
The intention for this second set of designs was to allow some local refinement of
the neural network metamodels in the vicinity of the design optimum, so that the
optimal design may be estimated more accurately. In this regard, the additional
designs function as a rudimentary variable-fidelity model.
7.4 SPLASH analysis

SPLASH is the potential flow code selected by the Alinghi design team for its CFD
analysis work. Execution times are reasonable, with a 200 point matrix of parameter
variations taking approximately 12 hours on a single CPU. This performance is
approximately two orders of magnitude faster than an equivalent RANS code.
SPLASH executes a matrix of test runs covering various combinations of hull

velocity, heel, yaw, rudder angle and trim tab angle. This matrix of test points is also
defined using the Uniform-Design quasi-random DOE method, in order to provide
uniform spacing of result data for the metamodel fitting process.
· 130 ·
Hydrodynamic data were calculated by SPLASH for each design, using the following
independent variables:
• VB,
• heel,
• yaw,
• rudder angle – set proportional to heel angle,
• trim tab angle – fixed angle for upwind, zero degrees downwind.
Dependent variables were:
• lift,
• residuary resistance,
• hydrodynamic heeling moment,
• wetted surface area of the canoe body,
• waterline length.
Note that although SPLASH is capable of calculating total resistance as well as

residuary resistance, in this case the Alinghi designers preferred to use the residuary
resistance value and to calculate the frictional resistance independently, using the
wetted surface area and waterline length provided by SPLASH.
7.5 Neural network metamodel creation

Neural network metamodels were created with the commercial neural network
package NeuroIntelligence, (Alyuda 2005), using the output data created by
SPLASH. Each of these metamodels used the following as independent variables:
BWL, CM, CP, LCB, topside flare, VB,, heel, yaw.
Metamodels were created for each family of designs, for each of the following values:
• lift,
• upright residuary resistance, RU ,
• heeled residuary resistance delta, ( Rϕ RU ),
• hydrodynamic heeling moment,

• wetted surface area of the canoe body in m2,
• waterline length, WL in m.
Metamodels were output from Neurointelligence as XML network definition files.
· 131 ·
7.6 VPP
VESPA was designed to work with any VPP that could be compiled as a DLL and
operated through a simple function call interface.
During its initial prototyping phase, VESPA was integrated with the SPAN VPP,
originally written by the author as part of the Maxsurf suite of design software.
However, SPAN was not specifically adapted to suit ACC yachts, and as a result was
not able to perform accurate ACC performance prediction.
For ACC optimisation, VESPA was interfaced to PAP, a VPP developed by a member
of the Alinghi design team, Manuel Ruiz de Elvira. PAP was originally developed for
the 2000 America’s Cup and was subsequently used for the design and analysis of
SUI-64, the winner of the 2003 America’s Cup. PAP is a reliable, well-validated
program used by the design team of the current holder of the America’s Cup, and
was seen as an ideal VPP on which to base an optimisation system.
VPPs typically use a regression model for hull lift and drag based on a series of tank
tests known as the Delft series (Gerritsma et al. 1981; Keuning and Sonnenberg
1999). Some VPPs, such as PAP, also have the capability to use lift and drag data for
a specific hull, derived from either tank test results or CFD calculations. This allows
more accurate performance prediction for a particular yacht. In order to work with
VESPA, this capability was extended to incorporate neural-network based
metamodels so that lift, drag and hydrodynamic heeling moment could be
estimated for variations to a parent ACC hull.
NeuroIntelligence, the neural network software used for the creation of metamodels
for VESPA, allows an XML file describing the architecture and weights of a trained
neural network to be written as output. PAP was modified to read these metamodel
definition XML files. Once read by PAP, the data contained in these files are used to
reconstruct the neural networks to allow hydrodynamic data to be derived for any
variation of the parent hull design required by VESPA. An example of an XML
neural-network definition is listed in Appendix 1.
In order to utilise this file format, it was necessary to write software that could read
the file and re-create the neural network from the definition contained in the XML
data. This neural network software module, named XML_NN, was written in such a
way that it could be compiled and linked directly into a C++ program, or
alternatively, the code could be compiled as a Microsoft Windows Dynamically
Linked Library (DLL).
Compiling as a DLL has several advantages. Firstly, it allows the software to be used
with a variety of other programs, even if these programs were written in different
languages. For example, the DLL was able to be called from within an Excel
spreadsheet, permitting extensive validation to take place utilising standard
· 132 ·
spreadsheet functions and charting facilities. The DLL was also able to be linked
with a Visual BASIC application and function as an integral part of that application.
This is the approach that was taken with the PAP VPP, which was modified to link to
the XML_NN DLL.
The XML_NN module was designed to be able to reproduce any neural network
architecture that could be created within NeuroIntelligence. XML_NN permits up
to five hidden layers to be specified, with up to 1000 neurons in each layer. Linear,
sigmoidal or hyperbolic tan activation functions may be selected on a layer by layer
basis. A listing of the header file for the XML_NN module, showing the class
structures, is provided in Appendix 2.
7.6.1 DLL link from VPP to optimiser
In addition to the PAP VPP calling the XML_NN module as a DLL, it was also
necessary for the VESPA optimiser to call the PAP VPP as a DLL. There are several
reasons for VESPA to use dynamically linked modules that are assembled at runtime.
VPPs are highly specialised programs, which have typically been developed over a
long period by particular design groups and America’s Cup teams. The internal
workings of these programs are proprietary and confidential, making it unlikely that
the author of such a VPP would be willing to provide source code or libraries to
enable a VPP to be compiled by an outside party.
By specifying an interface to a DLL, rather than requiring code to be compiled and

statically linked together, it is possible for each group providing software
components to keep their source code confidential.
This also facilitates updating or changing components, such as the VPP used,
without the need to recompile the whole system. DLLs may also be written in any
one of a number of languages. As long as the new VPP is compatible with the calling
interface of the DLL, it may be dynamically linked at runtime with no other
changes required.
PAP was modified to allow it to be compiled as a DLL and called from VESPA. A
listing of the header file for calling the PAP VPP module is provided in Appendix 3.
7.7 Race Modelling Program

The VESPA RMP was coded in C++ and is tightly integrated with VESPA’s genetic
algorithm code. The code covering the calculation of leg and race times from VMG
values for each discrete step is shown in Appendix 4. C++ code for the simulation of
races between two boats, including random time variation and trailing boat
penalties, is listed in Appendix 5.
· 133 ·
The VESPA race model consists of six legs, together totalling 20 nautical miles. Legs
are divided into 60 steps of approximately 100 metres, each step having its own
wind velocity and direction determined by sampling the wind distribution specified
for the event. VMG (or VMC if wallying is applied) is calculated for each step and the
total time for the leg accumulated.
One-hundred different sets of race conditions are created, each containing wind
velocity and direction for every step of the 6 race legs. Every boat in the population,
which typically consists of 25 individuals, has its leg times calculated for each of the
races. Once this is done, each boat is raced against every other boat (e.g. 25 × 24
races) for each of the 100 races, and each race is repeated 100 times with random
time variation included.
Each race consists of looking up the pre-calculated times for each boat for each leg
in turn. Times at the end of the leg are adjusted for random variation, port tack
penalty and trailing boat penalty. The adjusted times for each boat are then used as
the starting point for the following leg. This process is repeated until all legs are
completed, at which point the race is concluded and the winning boat has its
winning tally incremented.
The random time penalty applied at the end of each leg is normally distributed with
a standard deviation of six seconds. While it is possible that the amount of variation
added to the system should vary based on the wind velocities encountered during
each leg, with lighter winds having less variation and higher winds having more, the
analysis required to quantify these differences was beyond the scope of this research.
A sensitivity analysis, described in Section 8.8, showed that adding only a small
amount of random variation to the system had similar results to adding much
greater amounts of noise. It was therefore considered that using a random variation
appropriate to the mean wind velocity should not adversely affect results at the
extremes of the scale of wind velocities.
Once the random time variation has been incorporated, an analysis of starboard
tack advantage for windward marks is performed, with a penalty applied to port
tack boats that are neither clear ahead nor clear astern. This penalty eliminates any
overlap between the yachts, equivalent to a port tack yacht being forced to cross
behind a starboard tack yacht on an upwind leg.
This is followed by the application of a penalty for the trailing boat, the magnitude
of which is a function dependent on the time difference between the boats. This
penalty reduces in a linear fashion, with a maximum penalty of six seconds when
the two yachts are nearly overlapped, reducing to zero when the leading yacht is
ahead by twenty seconds.
The 25 × 24 round robin implies that yachts A and B race each other twice (A vs. B
and B vs. A). However, these races are not the same, as the advantage of the
· 134 ·
starboard entry to the starting line is allocated to the first boat in each pairing. For
two boats that are closely matched, this starboard end advantage may be the
deciding factor in the outcome of the race, and it is essential that this advantage is
equalised.
7.7.1 Wind model
VESPA uses an approximation of the wind velocity distribution defined by mean

wind velocity, standard deviation, and upper and lower wind limits. To verify that
this approach was a reasonable approximation to actual conditions, an analysis of
historical wind records for Valencia for the months of June and July was performed.
Figure 67 shows a histogram of wind velocities measured during June 2005,

between the hours of 13:00 and 16:00, together with a normal distribution curve
having a mean of 10 knots and a standard deviation of 2.5 knots. The historical
wind distribution is well approximated by the normal distribution in this case,
despite the short period during which data was collected.
0.16
0.14
0.12
Probability Density
0.10
0.08
0.06
0.04
0.02
0.00
5 7 9 11 13 15 17 19
Wind Velocity (knots)
Figure 67. Comparison of wind velocity distributions
An analysis of meteorological data spanning a ten-year period resulted in the use by

VESPA of a mean wind-strength of 11 knots, a standard deviation for the America’s
Cup event of 4 knots and a standard deviation for each race of 1.4 knots. A lower
limit of 6 knots and an upper limit of 20 knots were adopted to correspond with
wind strength limits used for actual racing.
7.7.2 Wallying
To apply wallying within VESPA, a standard deviation for the wind direction is
required, along with the extent to which wallying will be applied: 0%, 50% or
100%. The 100% value requires that the boat sails at the maximum wally value at
all times. This is not a realistic assumption, as it implies complete certainty as to the
future changes in wind direction.
· 135 ·
The 50% wallying setting can be thought of as using 100% wallying at the start of
the upwind leg, tapering linearly to 0% by the time the top mark is reached. This
case is more realistic, as it suggests that, at the start of the leg, the sailors will be
confident that a lift will revert to the mean before the top mark is reached, but that
this confidence in a shift reverting to the mean will progressively reduce as they get
closer to the layline for the mark.
7.8 Genetic Algorithm

A floating-point genetic algorithm was written in C++ using template-based classes.
A listing of the Evolver.h header file, showing the class structures for the genetic
algorithm code, is provided in Appendix 6.
The use of C++ templates allowed the GA code to be extremely flexible, separating
the representation of the genome from the mechanics of the evolution process. The
representation of the genome is defined by the Genome virtual base class and this
may be overridden to use a binary representation, a floating point representation or
even a mixture of the two. Similarly, genes are not restricted to being a binary or
floating point value, they may be entire data structures containing additional
information. For example, each gene could be set up store its own mutation rate,
and this could be evolved in parallel with the rest of the genome.
Mutation and recombination operators are also defined within the genome class,
making customisation of the GA straightforward and self-contained.
GA Operators
The following genetic operators were adopted, based on recommendations derived

from previous research by the author, (Mason et al. 2005):
• dual-point crossover;
• directional crossover;
• Gaussian distributed mutation;
• elitism;
• tournament selection.
Input variables to the GA were bounded to lie within the ranges specified for the
systematic series used for the fitting of the lift and drag metamodels. Values outside
these limits are unlikely to be predicted accurately.
Applying strict bounds to the input variables resulted in no infeasible designs being
encountered. However, future work that explores greater variable ranges will need to
consider the possibility that infeasible designs may be generated.
· 136 ·
7.9 User interface

VESPA was implemented as a Microsoft Windows program with multiple viewing
windows for hull and rig geometry, VPP results, polar performance graphs and
optimisation results. A typical layout showing these four main windows is shown in
Figure 68.
Figure 68. VESPA application
Controls were provided for VESPA’s optimisation parameters in a single dialog used
to commence the optimisation process. These controls include:
• genetic algorithm parameters, including the number of generations and the
size of the population;
• controls for including and defining exemplars where required;
• upper and lower limits for each design variable;
• design variable values for exemplars;
• parameters controlling the wind distribution used by the RMP;
• variables related to trailing boat penalties;
• Wallying options.
The dialog displaying these controls is shown in Figure 69.
· 137 ·
Figure 69. VESPA optimisation controls
Once these options have been specified, clicking the OK button commences an
optimisation run. Evaluation of a population of 25 yachts takes approximately 15-
30 minutes per generation on standard PC hardware.
Once completed, results of the optimisation run are presented in VESPA’s results
window, with members of the population sorted in descending order of fitness, as
shown in Figure 70.
Figure 70. VESPA results window
· 138 ·
Chapter 8 · Verification & Validation
8  Verification & Validation

The nature of optimisation is that it will ruthlessly exploit weak data. Rather than
minimising the function of interest, an optimisation procedure may simply locate
the point of maximum disparity between simulation and reality. As stated by Karl L.
Kirkman, one of the pioneers of racing yacht research:
“the first result, from my experience, of using optimization tools is to detect the
flaws in the algorithms.” (Oliver et al. 1987, p.261)
The accuracy of the results of an optimisation system such as VESPA is dependent

on the quality of both the source data and algorithms used. It is essential that the
performance of these components be verified and validated, both individually, as
well as when integrated into a complete system.
In the case of VESPA, this consisted of seven areas that required validation:
• parametric transformation;
• SPLASH results;
• neural network metamodel;
• accurate representation of the metamodel by the neural network DLL code;
• GA convergence and performance;
• VPP;
• RMP.
· 139 ·
8.1 Parametric transformation validation

The parametric transformation function was first checked to verify that the hulls
created matched all the design parameters specified. This was done by creating a
spreadsheet macro that automatically created several hundred design variations of a
parent model using the parametric transformation. Each of these was subsequently
analysed to verify that its hydrostatic and dimensional parameters matched the
values that had been used as inputs to the parametric transformation.
This testing resulted in some work being performed to improve the speed and
convergence of the parametric transformation algorithm. Additionally, the method
was refined to prevent introduction of hollows in the aft run of the hull that
appeared when CP was reduced or LCB moved forward.
8.2 SPLASH validation

SPLASH is a potential flow code that has been widely used by America’s Cup teams
for more than twenty years (Rosen et al. 1993; Rosen et al. 2000; DeBord et al.
2002 ). During this period, SPLASH has been extensively compared and validated
against tank test data and results from full scale, and as a result its capabilities and
limitations are well understood (Reichel et al. 1994; DeBord et al. 2002).
Consequently, no further effort was expended to validate SPLASH output for ACC
designs. Rather, results were examined critically and areas in which SPLASH was
considered to be giving potentially unreliable results were restricted in their range of
applicability. One example of this was the favouring by VESPA of designs with high
CP and aft LCB values. This trend was noted early during VESPA’s testing period, and
as the designs recommended as optimal had consistently higher CP and LCB values
than the best models from tank testing, the decision was made to restrict CP and LCB
ranges during optimisation.
This is not to say that the SPLASH data was in error, as it may have been that the
narrow BWL, high CM designs favoured by VESPA genuinely benefitted from higher
CP and LCB values. Rather, the limited time and tank testing resources available made
it impractical to attempt to validate this design direction. As a result, CP and LCB
values were restricted to known good values.
Final validation of the SPLASH predictions used by VESPA consisted of the

comparison of SPLASH data and tank data for the parent and optimised hull. In the
case where tank testing was performed, the deltas between the two hulls as predicted
by SPLASH and the tank corresponded well, as described in Section 9.5.3. However,
difficulties were noted in the ability of SPLASH to accurately predict the impact of
extreme values of some design parameters, such as combinations of high heel and
yaw angles.
· 140 ·
8.3 Metamodel training & validation

To simplify the neural network fitting task, output data from SPLASH were
modified to reduce the degree of non-linearity in the data. Lift, drag and
hydrodynamic heeling moment values were divided by 0.5 . While this did not
make the hydrodynamic heeling moment truly non-dimensional, removing the
relationship from the heeling moment data did assist with the approximation task.
The dataset was subsequently partitioned into three categories:

• Training Set. The discrepancies between the output values contained in the
training set and the values predicted by the neural network are used by the
training algorithm to determine changes to the network weights during neural
network training.
• Validation Set. The validation set is used as an independent measure of error.
Points in the validation set are not used for training the neural network and as
a result, may be used to indicate whether the network is generalising (i.e.
successfully predicting unknown points) or over-fitting (i.e. having improved
prediction accuracy for training-set data, but worse prediction accuracy for
validation set data). Over-fitting is generally indicative of an excess of neurons
in the hidden layer of the neural network.
• Test Set. The test set is comprised of data that has been put aside for the sole
purpose of providing an unbiased estimator of the prediction accuracy of the
trained neural networks. In this regard, the test set is not used for any purpose
that may result in it introducing bias into the training process.
Such bias may be an issue with the validation set, as it may be used to select
the best network at the completion of training. Consequently, the neural
network may be biased towards the contents of the validation set, even though
none of the points in the validation set were specifically used for training the
network.
For the VESPA metamodels, the independence of the test set was also
encouraged by the use of different test matrices for the test set hulls. Rather
than using the Uniform Design based sampling adopted for the training and
validation data sets, the test set design variables were generated randomly so
that their design variable values did not match any regular spacing.
Similarly, the test set hulls did not use a Uniform Design based matrix for their
array of hull velocities, heel and yaw angles. Instead, the sampling used the
Alinghi tank test matrix values, a full factorial matrix focussed around likely
combinations of hull velocity, heel and yaw.
Partitioning of the training set, validation set and test set approximated the ratio
50% : 25% : 25%.
· 141 ·
Training Algorithms
NeuroIntelligence contains several different training algorithms, including Back

Propagation, Quickprop, Conjugate Gradient, Quasi-Newton (QN), and
Levenberg-Marquardt (LM) methods. These algorithms differ significantly in their
performance characteristics and convergence properties, as shown in Table 8, which
lists the times taken to train a neural network to fit upright drag data to an RMS
error of 1.0 10 .
Table 8. Comparison of Neural Network Training Algorithms
Iterations Elapsed
Method Number of iterations / second time (s)
Backpropagation (Batch) unconverged after 100,000 n/a n/a

Backpropagation (Online) unconverged after 100,000 n/a n/a
Conjugate Gradient 6,827 8.34 818.6
Quickprop 10,573 65.11 162.4
Quasi-Newton 1,190 10.15 117.2
Levenberg-Marquardt 99 2.81 35.2
Despite the LM method significantly outperforming all other training algorithms in

this testing, it did not prove to be a suitable choice as the sole neural network
training method, as it appeared to stop prematurely before it had fully converged.
In a test where the LM method was used to train a neural network to the lowest
error achievable for heeled drag data, with training repeated 10 times to ensure the
best possible outcome, the results in row 1 of Table 9 were recorded.
When the QN method was used to train the network to the same error value
achieved by the LM method (Table 9, row 2), the elapsed time was more than thirty
times greater than that taken for a single run of the LM method.
Table 9. Convergence Properties of LM and QN Methods
Run No. of Iterations Elapsed

Method AAE
No. iterations /second time (s)
1 LM (best of 5) 53 0.92 57.5 2.3560
2 QN 6,075 3.93 1546.1 2.3556
3 LM (run 1) + QN (average of 3 runs) 5,000 3.90 1307.1 1.7838
4 QN (run 2) + QN (average of 3 runs) 5,000 3.88 1805.6 1.8959
5 QN 11,075 3.73 2975.6 1.9747
The results of runs 1 and 2 were both able to be improved significantly with further
training using the QN method. When each run was followed by an additional 5,000
· 142 ·
iterations of the QN method, the LM + QN sequence outperformed the QN + QN

sequence, as shown in rows 3 and 4. Finally, an unbroken run of 11,075 iterations
of the QN method gave slightly worse results than the QN + QN sequence, showing
that stopping and restarting the QN algorithm did not dramatically affect the
outcome and may have been beneficial.
These results are typical of those experienced during training of SPLASH data using
NeuroIntelligence. The NeuroIntelligence implementation of the LM method was
able to converge rapidly on a good solution, but stopped prematurely. The QN
method, on the other hand, was slow to find solutions of equivalent quality when
starting from scratch, but given the best result provided by the LM method, was able
to refine it significantly. This testing indicated that the best approach to neural
network training within NeuroIntelligence was to use repeated runs of the LM
method for initial training, followed by refinement using the QN method.
During initial metamodel fitting work, training was terminated if the validation
error was stable or worsening for several thousand iterations, as it was considered
that no further improvement was likely to be seen. While this appears to have been
a valid assumption for the standard backpropagation and Quickprop algorithms, the
quasi-Newton algorithm repeatedly showed that it was able to make significant
improvements after long periods of apparent stability in the validation set error
values.
An example of the benefits of extended training times is shown in Figure 71. In this
case, validation set error values worsen between 10,000 and 30,000 iterations. At
this point, typical early stopping criteria would halt the training process. If training
was allowed to continue, a very long period of no improvement follows until
240,000 iterations, at which point the validation error undergoes a dramatic
reduction of approximately 20%.
Validation Set
0.045
Training Set
0.040
Average Absolute Error
0.035
0.030
0.025
0 100,000 200,000 300,000 400,000 500,000
Iterations
Figure 71. Extended neural network training
· 143 ·
As a result of examples such as this, a policy of extended “training to exhaustion”

was adopted for VESPA metamodels, using the quasi-Newton algorithm for at least
250,000 iterations. In many cases this showed no benefit, with the training error
converging on a minimum after several thousand iterations. On other occasions,
extended training proved to be advantageous, with significant reductions in error
values occurring late in the training process.
Outliers
Initial neural network training of each data set focussed on outlier identification.
Outliers were common, due to both errors in SPLASH input parameters, as well as
the occurrence of non-converged points in SPLASH. Although NeuroIntelligence
has automatic outlier identification functions, it became apparent that this often
missed true outliers, while incorrectly classifying known good points.
It was found that the fastest way to filter outliers was to perform a preliminary
neural network fit using the Levenberg-Marquardt method. This preliminary fit
could be performed rapidly due to the exceptional performance of the LM method.
Initial fitting was followed by examination of the correlation plot for points that
were predicted poorly. An example of such a plot is illustrated in Figure 72. This
shows three small groups of outliers, each of which could be identified and removed
from the training data set.
Figure 72. Correlation plot of neural net prediction versus actual, showing outliers
· 144 ·
Outlier identification and removal is an iterative process. As outliers are removed,

the quality of each retrained neural network improves, allowing more outliers to be
identified.
Some outliers may be clearly associated with specific errors in input to SPLASH. In
these cases, the errors in the input values may be corrected and SPLASH re-run. For
example, in the case shown in Figure 72, the three groups of obvious outliers were
due to the incorrect specification of trim-tab angle in the model used for SPLASH,
resulting in anomalous lift and drag results. These points were successfully
recalculated using the correct trim-tab values and the correct results substituted into
the training data set.
Metamodel Training Process
Once outlier removal had taken place, the neural net training process within
NeuroIntelligence consisted of four steps:
• Architecture search. NeuroIntelligence is capable of performing an automated
search for the network architecture that results in the best performance. Upper
and lower limits are placed on the number of neurons to be incorporated in
the hidden layer of the network. The system will step through from least to
most, repeating neural network training a set number of times at each level
until complete, at which point the architecture with the best measure of merit
is selected.
This measure of merit may be training set error (not advisable due to the bias
it introduces to the architecture design), test set error (not advisable as it
makes the test set no longer truly independent), validation set error (best, but
introduces some bias), or a weighted sum of two or more of the three sets.
The adoption of the architecture with the smallest error is not recommended
as it may result in a network with too many neurons in the hidden layer and a
high risk of over-fitting. In this regard, Akaike’s criterion (Akaike 1974), which
weighs error values against network complexity in order to recommend a
compromise solution, was found to be a reliable predictor of the optimal
network architecture.
• Initial fitting was performed using the Levenberg-Marquardt algorithm with
between 5 and 10 retrains.
• Extended training was performed using the limited memory quasi-Newton
method for up to 250,000 iterations.
• Export of trained networks to an XML file for input to the VPP.
This process was repeated for each of the metamodels required by PAP, resulting in
the creation of six XML files for each family of designs derived from a particular
parent model.
· 145 ·
8.4 Neural network validation

Prior to fitting neural network metamodels, SPLASH output data was partitioned
into three separate groups. Of the data for the 24 hulls in the systematic series, 67%
were allocated to a training set and 33% to a validation set. An additional six hulls,
whose parameters were selected randomly, were analysed for use solely as an
independent test set. This allowed the prediction capability of the neural networks
to be tested to a high degree of confidence.
After extensive neural-network architecture search, followed by extensive training of

several alternative configurations, selection of the best neural network was
performed based on minimum test-set error.
Table 10 lists the average absolute error (AAE) and average relative error (ARE)
values for the neural network metamodels used for Parent Model 3, the final design
evaluated by VESPA. Analysis of training, validation and test sets showed that ARE
was less than 1% for all values.
Table 10. Neural Network Prediction Errors – Parent Model 3
Drag Drag Lift Heeling LWL Wetted

Upright Heeled (N) Moment (m) Surface
(N) (N) (N.m) (m2)
Average Absolute Error 2.676 9.144 79.583 82.616 0.020 0.040
Average Relative Error 0.29% 0.75% 0.77% 0.82% 0.09% 0.07%
Once optimisation runs had been performed and a likely area of the parameter
space isolated as a location for an optimal design, a further six to eight designs were
created and analysed in the vicinity of this design. In some cases, these additional
designs were used to extend the range of a design variable where the optimal design
had been found close to or at the upper or lower limit of the existing range for that
variable.
Once these additional designs were analysed using SPLASH, their hydrodynamic
data were incorporated into the original dataset. Neural networks were then
retrained, with a resulting improvement in the prediction accuracy in the area of the
optimal design.
8.5 Neural network DLL validation

The XML_NN module was tested extensively to ensure that it gave identical results
to the same neural network running within NeuroIntelligence. Testing verified that
results were identical in all cases, within the limits of double-precision, floating-
point rounding error.
· 146 ·
8.6 GA validation
The GA was validated using a variety of simple problems to ensure correct
functioning. The GA was also tested within the VESPA system using synthetic VPP
performance data to test the ability of VESPA to correctly find a known optimal
design. The synthetic VPP data was defined as:
1.5 T sin( TW ) cos( TW ) 0 TW 90

MG
2.0 T sin(180 TW ) cos( TW ) 90 TW 180
MG MG (1.0 |3.2 WL | 0.1) (1.0 |0.82 M|

(1.0 – |5.0 Flare| 0.05) (1.0 |0.58 (8.1)
P|
(1.0 |0.54 CB |)
This function resulted in simple polar curves having best upwind VMG at 45°and
best downwind VMG at 135°. Optimal performance occurred at design parameter
values of BWL = 3.2m, CM = 0.82, flare = 5°, CP = 0.58 and LCB = 0.54.
Test cases allowed either five design variables (BWL, CM, CP, LCB and Flare) or three
design variables (BWL, CM, and flare). Population size for this series of tests was set at
25.
The performance of the GA for this test function is shown in Table 11. This shows
that a close approximation to the correct optimum of the test function was reliably
found within 20 generations of the GA.
Table 11. Genetic Algorithm Performance for Synthetic VPP Data
No. of Free No. of BWL Flare

CM CP LCB ARE
Variables Generations (m) (deg.)
5 3.247 0.8153 0.5825 0.5397 5.0329 0.636%
10 3.216 0.8199 0.5821 0.5380 5.0217 0.305%
5 15 3.226 0.8208 0.5805 0.5398 5.0062 0.235%
20 3.199 0.8199 0.5802 0.5392 5.0023 0.053%
25 3.204 0.8200 0.5800 0.5392 4.9984 0.063%
5 3.191 0.8235 n/a n/a 4.8938 0.946%
10 3.204 0.8219 n/a n/a 5.0247 0.283%
3 15 3.207 0.8190 n/a n/a 5.0051 0.148%
20 3.200 0.8200 n/a n/a 5.0007 0.005%
25 3.201 0.8200 n/a n/a 4.9993 0.015%
Average Relative Error (ARE) values for the GA optimisation, using the test function
with three and five design variables, are graphed in Figure 73.
· 147 ·
1.00%
5 Free Variables
0.75% 3 Free Variables
Average Relative Error
0.50%
0.25%
0.00%
5 10 15 20 25
No. of Generations
Figure 73. Error values for different numbers of GA generations
Testing was also performed to determine the best population size for use by the GA.
If population size is increased while generation count is kept constant, a bias will be
created in favour of the larger populations, as the total number of evaluations
performed will be greater. To avoid such bias, this test maintained a constant
number of objective function evaluations, choosing combinations of population size
and number of generations that resulted in 192 objective function evaluations in all
cases. Results of the study are shown in Table 12 and Figure 74.
This testing showed that population sizes of between 12 and 32 gave similar results,
a figure that broadly agrees with the research cited in Section 6.4.5, which found
population sizes between 20 and 30 to be optimal. Smaller populations appear to
lack the diversity required for recombination operators to be effective, while larger
populations that use the same number of objective function evaluations have
insufficient generations for appreciable evolution to occur.
Table 12. Population Size Study Outcomes
Population No. of BWL Flare

CM CP LCB ARE
Size Generations (m) (deg.)
6 32 3.116 0.8195 0.4749 0.5399 5.0351 4.307%
8 24 3.189 0.8513 0.5709 0.5384 5.0093 1.241%
12 16 3.252 0.8212 0.5801 0.5403 5.0076 0.398%
16 12 3.193 0.8206 0.5820 0.5381 4.9673 0.328%
24 8 3.171 0.8187 0.5802 0.5408 4.9766 0.341%
32 6 3.196 0.8198 0.5815 0.5359 5.0392 0.389%
48 4 3.279 0.8165 0.5799 0.5419 5.0730 0.945%
64 3 3.165 0.8208 0.5794 0.5429 5.1989 1.166%
96 2 3.223 0.8257 0.5876 0.5337 4.7939 1.604%
· 148 ·
4%
Average Relative Error

3%
2%
1%
0%
0 20 40 60 80 100
Population Size
Figure 74. Average Relative Error values for variations in population size
As the test function employed was comparatively simple, it was considered that
populations smaller than 20 may be insufficient when more complex solution spaces
were encountered. As a result, a population size of 25 was adopted as a conservative
compromise between the conflicting needs of genetic diversity and low
computational cost.
8.7 VPP validation

Specific validation of the VPP was not performed, as PAP has been extensively tested
in the context of ACC yacht design during the past ten years. PAP has been regularly
compared to performance data from full scale testing and has been found to predict
ACC yacht performance reliably. PAP was used by the Alinghi team for the design of
SUI-64, winner of the 2003 America’s Cup.
The principal change made to PAP was the incorporation of the neural network
XML file definitions as the source for hydrodynamic lift and drag data.
Aerodynamic models and the algorithms used for balancing forces and moments
within the VPP were left unchanged. These changes required significant testing,
which revealed some convergence problems for PAP when the neural network
metamodels were not sufficiently smooth.
This problem was addressed in two ways. PAP’s solver was improved to make it less
sensitive to noise in the metamodels. More importantly, procedures for training
neural network metamodels were amended to avoid over-fitting of the source data.
The use of large test and validation sets, the adoption of extended training periods
and the selection of the smallest possible number of neurons in the hidden layers
were key components of this strategy.
· 149 ·
8.8 RMP validation

The VESPA RMP was coded and tested using standard software development and
testing methodologies. These included unit testing at the subroutine level, using a
set of test data with software stubs to uncoded subroutines, followed by system level
testing using synthetic data.
The VESPA RMP was tested using the same synthetic VPP data previously used to
validate the GA optimiser. Statistical analysis of actual race data was used to quantify
some of the values used for the RMP, such as trailing boat penalties applied at
rounding marks.
Random time variation
Section 5.2.4 describes the analysis of real-world test data to determine the random
time variation used by the VESPA RMP. This approach carries with it the risk of bias
due to the small size of the available sample.
To determine the impact of such a bias on optimisation outcomes, a sensitivity

analysis was performed using the race modelling capabilities of VESPA. For this
analysis, each race was repeated 100 times with a small random time penalty added
to each boat’s time for each leg. The SD of this variation was incremented from zero
to twelve seconds in three-second steps, with the results shown in Table 13.
Table 13. Change in Optimal BWL with the Incorporation of Random Noise
Optimal BWL (m)
Random Variation Wind velocity SD Wind velocity SD

(s) 1.0 knots 1.5 knots
0 3.116 3.170
3 2.999 2.997
6 2.997 2.997
9 2.999 n/a
12 2.997 n/a
Results of this analysis showed that the introduction of very small random variations
into the tournament model resulted in an immediate change in the best parameter
values found by VESPA, but that further increases to the magnitude of the random
variation had little additional effect on the results.
This was an encouraging result as it shows that the outcome was not highly sensitive
to the amount of random variation provided some variation occurred. This allowed
a conservative value for the SD of the variation to be chosen in the knowledge that it
was unlikely to have a detrimental effect on the outcome.
· 150 ·
8.9 VESPA system testing
8.9.1 Objectives and methodology
To evaluate the functioning of the entire VESPA system, a series of validation tests
were performed using an actual ACC design. The measure of success for these tests
was an evaluation of the recommended optimal designs using SPLASH and PAP,
followed by comparison with similar data for the parent model. For VESPA to be
successful in optimising the design, significant improvements in performance over
the parent model should be evident in these comparisons.
Using a parent model supplied by Alinghi, twenty-five variations were generated

using a uniform design sequence within the parameter limits shown in Table 14, to
give the hull shapes shown in Figure 75. These limits were chosen based on ranges
considered by Alinghi design team members as likely to contain optima.
Table 14. Parameter Variation Limits – Validation Model
BWL Flare
CM CP LCB
(m) (deg.)
Minimum 2.9 0 0.805 0.555 0.5275
Maximum 3.4 8 0.855 0.575 0.5475
Figure 75. Family of 25 hull design variations
In addition, five hull designs were created using random values of the above
parameters within the ranges shown. These additional hulls were created solely for
independent validation of the neural network metamodels.
· 151 ·
Appendages for the transformed hulls were identical to those of the parent model,
although keel and bulb were moved longitudinally to maintain the alignment of the
LCG of the yacht with its LCB. In addition, the longitudinal location of the rudder
was moved to maintain a constant distance between its trailing edge and the aft end
of each hull’s waterline.
These thirty hull designs were exported as IGES NURBS surface files and analysed
using the SPLASH potential flow code. Once SPLASH analysis was completed,
results for all hulls were collated into a single file and used as input to the
NeuroIntelligence neural network software. Of the data for the initial twenty-five
hulls, 67% were allocated at random to the training set, with the remaining 33%
allocated to the validation set. Data for the five randomly generated hulls were
allocated to an independent test set.
Neural network metamodels were trained for each of the following variables –
• lift;
• drag upright – residuary resistance for the zero heel, zero yaw case;
• drag delta – additional drag due to heel and yaw;
• hydrodynamic heeling moment;
• waterline length;
• wetted surface area.
Lift, drag and hydrodynamic heeling moment values were partially linearised by
dividing each by to remove the component. For drag, this gives the drag
area , the area of a flat plate, held normal to the direction of flow with a 1,
having equivalent resistance.
Converting the lift, drag and hydrodynamic heeling moment data in this way
reduces the degree of non-linearity in the data and allowed more accurate fitting of
the neural network metamodels. If this linearisation of the data is not performed,
least-squares fitting of the data is dominated by the errors for higher hull velocities,
with the fitting of low speed data suffering as a result.
8.9.2 Optimisation runs
Three optimisation runs were performed with VESPA using two different wind
velocities and wind velocity standard deviations. The purpose of these runs was to
investigate the effect of different wind velocity parameters on the hull design. While
a correlation between BWL and wind velocity may be a reasonable expectation, the
effect of changes to the variance of the wind velocity are not as easily predicted.
· 152 ·
Two different mean wind velocities were chosen, along with two different standard
deviation values for wind strength, in order to establish trends for the two variables.
Two different wind velocity standard deviation (SD) values are used; one SD for the
event, a total of 100 races, and one SD for each race. This is necessary because of the
different timescales involved. An event may take place over a period of several weeks,
while a race will take place over a period of 2-3 hours. The SD of the wind velocity
distribution for each of these two periods will differ significantly.
8.9.3 Results and validation
Results of the three optimisation runs are summarised in Table 15.
Table 15. Results Summary – Validation Model
Mean wind Event SD Race SD BWL Flare Number of

CM
velocity (knts) (knts) (knts) (m) (deg.) Generations
Run 1 16 4 1 3.376 0.911 0.812 10
Run 2 11 4 1 3.189 0.076 0.805 15
Run 3 11 1 1 3.023 0.105 0.805 20
For Run 1, VESPA searched for optimum values for BWL, Flare, CM , LCB and CP. The
GA was run for 10 generations. However, it was clear that in this case the GA did
not fully converge, as there was significant variation of parameter values in the
population.
BWL for Run 1 was higher than expected. Flare was low, but with significant
variation within the population. This was initially interpreted as incomplete
convergence on an optimal value. However, later runs also showed anomalous
scatter.
Optimisation Run 2 used a mean wind of 11 knots and an event SD of 4 knots. For
this and subsequent runs the LCB was fixed at 53.5% and the CP fixed at 0.57. The
purpose of fixing these values was to simplify the search process by reducing the
number of dimensions of the parameter space. This allowed the optimal values for
other parameters to be found in fewer iterations. This optimisation was run for 15
generations of the GA.
Optimal BWL reduced for this run, as did CM. Flare approached zero for the most
successful boats, although once again, variance for this parameter was high. CM was
at the lower limit of the test matrix range.
Optimisation Run 3 used a mean wind of 11 knots and an event wind strength SD
of 1 knot. Although this is not a realistic value for an event wind strength SD, it was
used to establish the trend in BWL and CM parameters as SD is reduced.
· 153 ·
This run was performed with fixed LCB and CP, and was run for 20 generations of
the GA to encourage convergence. Once again, the optimal BWL reduced and the CM
was at the lower limit. Variance for BWL and flare was small, indicating good
convergence. However, it is likely that lower CM values would have resulted had
VESPA been free to use them.
The trend towards lower BWL at lower wind velocities was not unexpected, as lower
viscous and residuary resistance resulting from lower beam, even at the cost of
reduced stability, is a reasonable trade-off for light weather. The unexpected
outcome of these initial test runs was the low CM values predicted. Based on the
previous experience of Alinghi’s design team, this appeared to be an anomalous
result.
On the basis of the trends seen in the three optimisation runs, and in order to
explore the low BWL, low CM parameter range, four variations of the parent hull were
created using two BWL values and two CM values, as shown in Table 16.
Table 16. Test Design Parameters – Validation Model
BWL Flare
CM
(m) (deg.)
VP_P2A 3.075 0.815 0
VP_P2B 3.075 0.805 0
VP_P2A2 2.975 0.815 0
VP_P2B2 2.975 0.805 0
The four test designs were analysed directly using SPLASH and their performances
compared with their parent model, using the performance comparison charts shown
in Figure 76.
These charts indicate relative performance of a yacht or yachts against a baseline

boat, whose performance is shown as a horizontal line along the abscissa. A yacht
having better performance than the baseline boat for a particular wind velocity,
measured in metres per minute, will have a value plotted below the abscissa.
Conversely, performance inferior to the baseline boat will show as points plotted
above the abscissa.
The four hulls tested showed good performance downwind compared with the
parent design. Upwind performance, although superior in less than 10 knots of
wind, was generally worse than the parent design above 12 knots of wind. Of the
four designs, the two narrow boats were fastest downwind and slowest upwind. For
equivalent beam, the boats with the higher CM value were superior upwind and
equal to, or better, downwind. The boat with possibly the best overall performance,
VP_P2A, had the highest BWL and CM values.
· 154 ·
Upwind
1
Advantage/disadvantage (m/minute)
0.5
-0.5
VP Parent
VP_P2A
-1
VP_P2B
-1.5 VP_P2A2
VP_P2B2
-2
6 8 10 12 14 16 18 20
Wind speed (knots)
Downwind
0
-0.5
-1
-1.5
VP Parent
VP_P2A
-2
VP_P2B
-2.5 VP_P2A2
VP_P2B2
-3
6 8 10 12 14 16 18 20
Wind speed (knots)
Average Upwind + Downwind

0.75
0.25
-0.25
-0.75
VP Parent
VP_P2A
-1.25
VP_P2B
-1.75 VP_P2A2
VP_P2B2
-2.25
6 8 10 12 14 16 18 20
Wind speed (knots)
Figure 76. Performance comparison plots of test models versus validation model
· 155 ·
Although these results demonstrated that the four test designs improved on the
performance of the parent model, they did not confirm the trend to low CM
suggested by VESPA. It was clear that VESPA had failed to correctly determine the
optimal set of values for the design variables under examination.
The failure of VESPA to perform as expected was a cause for concern and significant
time was spent determining the source of the problem. The initial focus of these
investigations was the numerical stability of PAP, the VPP used by VESPA.
PAP was originally an interactive program that calculated a small number of sailing
performance values, which were then used to fit polar curves. When such a program
is used as a “black box” solver for an optimisation loop, reliable operation is essential.
Instead of a small number of data points being generated, a very large number of
cases, potentially in the millions, must be accurately calculated. Numerical
instability may result in a small percentage of cases failing, resulting in erratic
behaviour by the optimisation procedure, or convergence to a false optimum.
Analysis of PAP’s error logs revealed that for a small percentage of cases, PAP failed
to converge, resulting in the calculation of an incorrect hull velocity. Although
measures were put in place to trap errors and assist PAP’s solver to converge on a
correct answer, convergence problems remained. After further investigation, it
became apparent that these were caused by small regions of the hydrodynamic data
that were poorly fitted by the neural network metamodels.
Examination of the neural network metamodels was performed by comparing

metamodel predictions for the test designs with actual SPLASH outputs. These
comparisons showed significant discrepancies between predicted and actual results
for drag at high heel angles. To assist in the visualisation of these errors an Excel
spreadsheet was created that allowed slices to be defined through the neural network
hypersurfaces for any parameter combination (Figure 77). These slices were
displayed as charts as shown in Figure 78, which compares neural network
predictions with SPLASH results for 15 degrees of heel and 1.5 degrees of yaw.
Figure 77. Validation spreadsheet controls
· 156 ·
Figure 78. Metamodel validation graphs
· 157 ·
Examination of these graphs revealed that an error existed in the metamodel for
heeled drag, which resulted in resistance decreasing significantly at high heel angles
(Figure 79). This had the effect of rewarding yachts with lower stability, for example
those with low BWL and CM values, resulting in these hull forms dominating the
optimisation process.
Figure 79. Drag errors for 11 knots velocity, 1.5 degrees of yaw
8.9.4 Validation model conclusions
No firm conclusions regarding design optima could be drawn from VESPA testing
of the validation model, due to metamodel fitting errors. As a result of these
difficulties, a significant effort was put into determining the most effective training
process for the neural networks metamodels. Systematic testing of different training
algorithms, network architectures and training periods was performed, together
with the development of procedures for effectively testing and validating the neural
network metamodels.
The SPLASH analysis performed on the four test designs did provide support for the
following observations:
• Reduced BWL improved downwind performance in all wind velocities, with the
greatest gains occurring in light winds.
• Reduced BWL resulted in improved upwind performance in light winds, but
was responsible for poorer upwind performance in greater than 12 knots of
wind.
• Lower CM values did not appear beneficial in any wind velocity, upwind or
downwind.
· 158 ·
Chapter 9 · Results & Discussion
9  Results & Discussion

Optimisation runs using VESPA took place between June and August 2006, using
three different parent hull models supplied by the Alinghi design team.
During this period, Alinghi also continued with its conventional design
development process. This consisted of hull shape experimentation by lead yacht
designer Rolf Vrolijk, CFD analysis using SPLASH and other CFD codes, simulation
of sailing performance using VPP software, followed by one-third scale tank testing
of the most promising designs.
Over the period during which VESPA was first run and progressively refined, the
Alinghi design team independently explored many promising design directions.
Some of these proved beneficial and were incorporated in the final design of SUI-
100, the yacht that went on to win the 2007 America’s Cup, while others proved to
be dead ends and were rejected.
Design avenues that were not successful had often shown benefits during CFD
analysis, which were subsequently not confirmed by tank testing. This is one of the
realities of modern yacht design and analysis; the numerical tools used are imperfect
and can easily result in false leads being followed, unless results are rigorously
validated.
· 159 ·
As a result of this ongoing design development process, the parent model used by
VESPA for optimisation changed several times. In some cases, the parent models
used were later found to be retrograde steps; at other times the parent, when tested
in the towing tank, was shown to be the current best of breed.
As the development process continued and the time remaining before delivery of
final hull lines for SUI-100 reduced, the time available for each cycle of VESPA
testing shortened. While parent model 1 was used for VESPA optimisation work for
two months, the time available for VESPA analysis of parent model 3 was less than
10 days. During this period, the process for generating the systematic series,
performing SPLASH analysis and creating fully trained neural networks became
more streamlined and automated, with significantly less manual intervention
required.
As testing progressed and refinements such as trailing boat penalties were added to
the VESPA RMP, it became apparent that the optimisation problem was
significantly more complex than first thought, with implications for both the design
of the optimisation algorithm as well as for the interpretation and application of its
results. These issues and their implications are discussed in Section 9.1.
9.1 Consequences of trailing boat penalties

The VESPA RMP models interactions between yachts at turning marks via the
application of penalty functions applied to the trailing boat, as described in Section
5.2.4. These penalty functions were still being developed at the time testing of
parent model 1 commenced, and consequently, no trailing boat penalties were
applied for the initial runs of VESPA using parent model 1. For later runs using
parent model 1, trailing boat penalties were utilised without the port tack penalty,
which was first implemented for use with parent model 2.
As penalties are applied at windward marks but not at leeward marks, their effect is
to favour boats that perform well upwind, at the expense of boats whose
performance profile is biased towards downwind work.
When no trailing boat penalties are used, the VESPA solution space often appears
bimodal. Yachts whose performance is biased to upwind sailing in preference to
downwind sailing may have similar elapsed times for an upwind/downwind
racecourse as yachts whose performance is biased towards downwind performance.
Once trailing boat penalties are applied this bimodality disappears, as yachts that are
fast downwind are unable to recover their upwind losses. Unfortunately, the removal
of this bimodality does not imply a simpler solution, with the effect of trailing boat
penalties on the ideal design parameters adding significant complexity to the
problem.
· 160 ·
During testing with trailing boat penalties, it became apparent that VESPA could
produce different results for two similar runs. Optimal hulls produced would be
similar in most regards, but could vary in BWL over a range of approximately 150-
200mm. It was clear that the optimum was not stationary; rather, it shifted based
on the makeup and diversity of the population at the time. If the population was
not diverse (i.e. design variables for most of the boats were similar), the optimum
could move around, whereas when the population was made up of a variety of
different designs, the optimum would tend to stabilise at high BWL values.
Testing was performed to investigate this problem. In the first set of tests, VESPA
was run for 25 generations, with a population size of 20, using parent model 1.
Results are plotted in Figure 80.
3.3
3.25
3.2
3.15
BWL
3.1
3.05
3 TBP
No TBP
2.95
0 5 10 15 20 25
Generations
Figure 80. Investigation into the effects of Trailing Boat Penalties (TBP)
Without trailing boat penalties, the optimal BWL value stabilizes after only a few
generations at approximately 3.05m. When trailing boat penalties are introduced,
the BWL increases to more than 3.25m and appear to stabilise, before dropping
suddenly to less than 3.1m. This instability in the solution did not always occur,
and on many occasions, the design with the highest fitness simply stabilised at a
high BWL value. However, with trailing boat penalties in use VESPA could not be
relied on to give reproducible results.
Nash Equilibria and mixed strategy games
To explain this behaviour, it is necessary to introduce the game-theory concept of a

pure-strategy Nash equilibrium. Nash (1950) proposed that in a two-player game, a
pure strategy existed if one player could adopt a strategy that did not need to be
changed, regardless of how an opponent’s strategy changed. In yacht design, this is
equivalent to finding the design parameters for the fastest boat for an upwind leg for
a fixed wind velocity. Regardless of how your opponent changes their design, it will
not be faster than yours. If a pure-strategy Nash equilibrium exists, it is possible to
locate a single optimum that cannot be improved on for a specific set of conditions.
· 161 ·
The introduction of a penalty for the trailing boat at the first mark, say 1.5-boat
lengths, changes the situation. As an example, assume a boat A, which has the best
aggregate of upwind and downwind performance when no trailing boat penalty
exists. A boat that has a slightly greater BWL (boat B) may be 2 boat lengths faster
upwind and 3 boat lengths slower downwind, and therefore slower by 1 length for
an upwind/downwind course. However, if boat A incurs a 1.5 length penalty at the
first mark, boat B will win the race by ½ a length. If both competitors in a match
race choose boat B for their design, neither will gain an advantage.
If a third boat, C, is slightly wider again, to the extent that it is 2 lengths faster than
B upwind and 3 lengths slower downwind, B will incur the 1.5 length penalty at
the first mark. In this case, boat C will win by ½ a length.
B beats A and C beats B. Is C therefore the best design to choose? Unfortunately,

when C races A, it loses. Boat C is 4 lengths faster than A upwind and picks up the
1.5 length upwind bonus, giving it a 5.5 length upwind advantage. However, it is 6
lengths slower downwind, so it loses overall by half a boat length. In this case, if
each competitor could choose only from the designs of boats A, B or C, there is no
strategy that guarantees success. In a round robin tournament between the three,
each boat would come away with one win and one loss.
This type of problem is considered to have a mixed-strategy equilibrium – there is

no single strategy that guarantees success, and there are multiple strategies that have
an equal chance of success. This is illustrated by the following diagrams, which use
idealised performance values for hypothetical boats. Figure 81 shows the change in
upwind and downwind performance as BWL is varied. As BWL increases, the upwind
performance improves while the downwind performance deteriorates. The fastest
boat overall is the one with best average of upwind and downwind performance,
shown by the peak in the upwind + downwind average curve.
Figure 81. Optimal beam for a simple yacht performance model
· 162 ·
The situation changes when a trailing boat penalty is introduced. In this case, the
wider boat in any match will have the performance shown by the upwind +
downwind average curve. However, the narrower boat of the pair will be penalized
on the upwind leg and its aggregate performance will be that of the curve running
parallel to and below the unpenalised performance curve. In this situation, a mixed-
strategy outcome occurs, as illustrated by Figure 82.
Figure 82. Effect of trailing boat penalties on race performance
In the first match, B versus A, boat A incurs a trailing boat penalty and B wins. In
the second match, B incurs the penalty and C wins. In the third match C is the
faster boat upwind, but A’s aggregate performance, even with the penalty included,
is superior to that of C. No strategy is superior, and all boats will score the same in a
round robin tournament.
Several mixed strategy games are similar to this problem. If we rename boat A
“Rock”, boat B “Paper” and boat C “Scissors” the similarities become obvious.
• Paper beats Rock.
• Scissors beat Paper.
• Rock beats Scissors.
In a game of rock, paper, scissors, no single strategy is superior because the hands
are revealed simultaneously. However, if a short delay occurs between the revealing
of the hands, the later player will always be able to choose a winning hand. In the
America’s Cup, the situation is similar when two teams are choosing boats for the
final round of the event. If both boats must be nominated simultaneously, neither
team can guarantee a winning strategy. If one team is forced to nominate their
design in advance of the other, and the later team is able to determine the design
parameters and performance profile of the first team’s boat, a boat that is slightly
superior upwind performance should be chosen.
· 163 ·
Note that it does not necessarily pay to be significantly faster upwind – a boat that
was halfway between boat A and boat B, so that it was only 1 boat length faster
upwind and 1.5 boat lengths slower downwind, would win the race by 1 length,
whereas boat A beat boat B by only ½ a length. Therefore, the strategy for an
America’s Cup defender would appear to be to choose a boat that is similar to the
challenger’s boat, but slightly biased towards upwind work.
For the challenger the situation is slightly different. Rather than having to compete
against one boat, challengers are each required to compete against ten other teams
for the right to challenge for the cup. This makes the problem of boat selection a
different one to the problem faced by the defender. Each challenger needs to select a
boat that will have the highest win/loss ratio when matched against a diverse range
of competing boats. In this case, there is an optimum, although it is dependent on
the makeup of the fleet.
It is possible to perform a simple Monte Carlo simulation to show how the

probability of winning varies with beam against a fleet of opponents in a round
robin format. Figure 83 shows the outcome of an analysis using a simplified yacht
performance model, in which 50 boats of different beams were matched against one
another. It can be seen that the curve of average win/loss ratio peaks to the right of
the BWL of the fastest boat, but drops dramatically once past the peak. It is clear that
the effect of applying a trailing boat penalty at the windward mark is to bias the
optimal design in the direction of greater upwind performance via increased BWL.
Figure 83. Win/loss ratio for a range of BWL values, utilising trailing boat penalties
The above examples are for a simplified and idealized performance profile for one
wind velocity. As the VESPA RMP uses a wind velocity distribution rather than a
single wind velocity, it was expected that the clear peak and sharp decline shown in
the curve above would be blurred substantially, but that a shift of the optimal beam
to the right and a sharp decline once past the optimum should be apparent.
· 164 ·
To test this hypothesis, a tournament model using the VESPA RMP was run for a
test design whose beam was varied over a range of 0.35 m. One tournament was
performed with trailing boat penalties turned on and a second with trailing boat
penalties turned off. The results are shown in Figure 84.
Figure 84. Effect of trailing boat penalties
Without trailing boat penalties, the curve of win/loss ratio is almost flat from about
3.070m to 3.110m, but declines quickly at higher beams. With trailing boat
penalties included, the range of optimal beam is broader and is shifted to the right.
It also starts to decline rapidly at beams above 3.160m. This area of rapid decline is
where the rock, paper, scissors effect applies, and sets the limit on how wide a boat
can be made before it becomes rapidly uncompetitive.
The introduction of trailing boat penalties resulted in the optimal solution changing
from having one or two stationary optima, to one defined by a limit cycle with no
stationary optimum. This required that the GA be modified to limit any cycling of
the solution to the minimum range, as it was possible for a co-evolutionary arms
race to result in design variables such as BWL drifting away from realistic values.
9.1.1 Genetic algorithm modification
One effect of the use of trailing boat penalties was the occurrence of co-evolutionary
arms races in the population, resulting in a solution that increased monotonically in
BWL, or cycled over a range of BWL values, rather than settling on a single
equilibrium solution. In order to prevent this occurring, two related measures were
implemented within VESPA, both based on the “hall-of-fame” approach of Rosin
and Belew (1996).
· 165 ·
In addition to using simple elitism, which retains the best individual from the
previous generation, a modified elitism method was implemented within VESPA.
Using this method, the individual retained from the previous generation is the one
that had the fastest average time when sailing all of the courses alone. Although this
individual may not be equivalent to the boat that wins the most races, it will be of
sufficient quality to provide an objective performance benchmark for other
members of the population, and may prevent cycling or drifting of best fitness
values due to disengagement.
Exemplars were also evaluated as a method for avoiding cycling of the solution and
allowing the GA to settle on a single optimal design. Exemplars have the advantage
that the design parameters for a known or expected competitor can be specified,
allowing VESPA to determine the optimal design to use against that competitor.
Test runs were performed using the same parent model and parameters as used for
the testing shown in Figure 80, but with modified elitism and a single exemplar.
The single exemplar used BWL = 3.05m. Results are shown in Figure 85.
3.3
3.25
3.2
3.15
BWL
3.1
3.05
TBP
3 No TBP
TBP + Hall of Fame
2.95
0 5 10 15 20 25 30
Generations
Figure 85. Investigation into the effects of exemplars and/or Hall of Fame
Although these measures may assist VESPA to converge on a single optimal solution,
this design is not necessarily an optimal solution against all opponents. The
implication is that, rather than a single optimal design existing for a given set of
weather conditions, the optimal design must instead be considered within a game-
theoretic framework and needs to be selected based on the knowledge or expectation
of the design parameters of the opposing yacht or yachts.
· 166 ·
9.2 parent model 1, May 28, 2006

Design optimisation of high quality design candidates was commenced during May
2006. A family of twenty-five models was generated from the parent using the
parameter ranges show in Table 17.
Table 17. Parameter Variation Limits – Parent Model 1
BWL Flare
CM CP LCB
(m) (deg.)
Minimum 2.9 0 0.80 0.575 0.53
Maximum 3.4 8 0.85 0.585 0.54
In addition, seven extra hulls were generated with random parameters within the
same ranges. These seven hulls were used solely for testing of the results of the
neural network training process.
The thirty-two hulls, along with the original parent design, were analysed by
SPLASH and the results collated into a single file. This data file was then used to
train neural networks for upright residuary resistance, heeled drag delta, lift,
hydrodynamic heeling moment, LWL and wetted surface area.
The results of the training were tested and validated by comparing with the original
training and testing data, both in tabular form and using the validation spreadsheet.
The results of this validation were excellent, with average relative error values less
than 1% in all cases. Fitting of the data was visually checked using the validation
spreadsheet, in order to ensure that there were no oscillations or areas of poor fit in
the neural network metamodels.
Once neural networks metamodels had been trained, several runs were performed to
determine optimal design parameters for different wind velocities and wind velocity
variance, shown as runs 1 to 7 in Table 18. CP and LCB were fixed for these runs to
values in the centre of the ranges analysed by SPLASH.
Table 18. Parent Model 1 Results
Mean wind Event Race BWL Flare Trailing boat

velocity SD SD CM Wally
(m) (deg.) penalty (s)
(knts) (knts) (knts)
Run 1 9 4 1.4 3.040 0 0.85 0 no
Run 2 11 4 1.4 3.050 <0.1 0.85 0 no
Run 3 11 4 1.4 3.043 0 0.85 0 no
Run 4 13 4 1.4 3.040 <0.1 0.85 0 no
Run 5 15 4 1.4 3.045 <0.1 0.85 0 no
Run 6 11 8 1.4 3.063 0 0.85 0 no
Run 7 11 8 2 3.029 7.81 0.85 0 no
· 167 ·
VESPA runs were performed for mean wind velocities of 9, 11, 13 and 15 knots to
determine how the optimal midship section parameters varied. A second run for 11
knots was performed using a different starting population, in order to test the
reproducibility of the system. This repeated run achieved virtually identical results
to the original.
Surprisingly, there was only a small variation in beam for these changes in wind
velocity, with the optimal BWL varying by only 0.010m across the range of mean
wind velocities.
For the conditions similar to those expected in Valencia at the time of the 2007
America’s Cup (mean wind 11 knots, S.D. 4 knots for the month of June, 1.4 knots
S.D. per race) VESPA predicted a boat with a BWL of between 3.040 and 3.050m,
zero flare and a CM of 0.85.
Two runs (numbers 6 and 7) with increased wind velocity variance were performed
with unexpected results. The wind velocity SD for the event was increased from 4 to
8 knots and this resulted in a small increase in BWL of between 0.013 – 0.020m.
However, when the SD for each race was also increased from 1.4 knots to 2 knots,
the beam of the optimal boat reduced significantly to 3.029 m while flare increased
to 7.8 degrees, very close to the upper limit tested for this parent model.
This result again raised the possibility of a bimodal solution, with designs having
moderate beam and zero flare dominant in steady conditions, while boats having
narrower beam and significant amounts of flare showed promising results in more
variable conditions.
One possible explanation for this is that a boat that is required to perform in a wide
range of wind conditions may use a narrow BWL to reduce drag, giving an advantage
in light wind and on downwind legs, while regaining the stability needed for
upwind sailing and heavier conditions by the addition of flare to the topsides.
As a result of these anomalous results, two test designs with different beam and flare
characteristics were generated for SPLASH analysis and comparison with the original
parent model. The design parameters and hull shapes for these comparison models
are shown in Table 19 and Figure 86.
Table 19. Test Model Parameters – Parent Model 1
BWL Flare
CM
(m) (deg.)
R1 3.050 0.85 0
R2 3.030 0.85 7.8
R1W 3.168 0.85 0
· 168 ·
Figure 86. Test hulls R1 (top) and R2 (bottom)
These boats were run through SPLASH and compared to the parent design (Figure
87). Comparisons showed that boat R1 had a speed advantage over the base boat in
almost all wind velocities, both upwind and downwind.
This testing also demonstrated that a bimodal solution was possible. Boat R2 had a
similar aggregate upwind/downwind performance to R1. However, R2 achieved this
by combining comparatively poor upwind performance with superior downwind
speed. While the R2 design may have been at a significant disadvantage in actual
racing conditions, which are likely to benefit boats having an upwind speed
advantage, it was clear from these results that determining a single optimal design
for a specific set of conditions was not as simple as first thought.
Subsequent to the positive outcomes shown by test design R1, members of the
Alinghi design team expressed concern that R1 would be at a disadvantage in actual
racing compared to a boat with greater beam. As a result, a wider version of R1 was
generated and tested in SPLASH. This hull is designated R1W in Figure 87.
Although R1W was clearly superior upwind, it was at a disadvantage downwind
against all other hulls, and its averaged upwind/downwind results were generally
inferior to R1.
· 169 ·
Upwind
0.5
-0.5
VP Parent
R1
-1
R1W
R2
-1.5
6 8 10 12 14 16 18 20
Wind speed (knots)
Downwind
0.5
-0.5
VP Parent
R1
-1
R1W
R2
-1.5
6 8 10 12 14 16 18 20
Wind speed (knots)
Average Upwind/Downwind
0.5
-0.5
VP Parent
R1
-1
R1W
R2
-1.5
6 8 10 12 14 16 18 20
Wind speed (knots)
Figure 87. Performance comparison plots of test models versus parent model 1
· 170 ·
9.2.1 Trailing boat penalties
The optimisation runs detailed above were performed with no trailing boat
penalties or attempts to bias upwind versus downwind performance. To investigate
the effects of trailing boat penalties, several runs were performed with a simple
trailing boat penalty, with the results listed as runs 8, 9 and 10 of Table 20.
Table 20. Results With Trailing Boat Penalties Included – Parent Model 1

(m) (deg.) penalty
Run 8 9 4 1.4 3.091 <0.1 0.85 6 no
Run 9 11 4 1.4 3.100 <0.1 0.85 6 no
Run 10 13 4 1.4 3.105 <0.1 0.85 6 no
Run 11 11 4 1.4 3.100 <0.1 0.87 6 no
Run 12 11 4 1.4 3.111 <0.1 0.87 6 yes
Compared to the tests listed in Table 18, trailing boat penalties increased the
optimal beam by 50-60mm and resulted in the emergence of a correlation between
BWL and wind velocity.
CM parameter range extension
One concern regarding these results was that the CM values of the predicted boats
were at the upper limit (0.85) of the range tested. In order to check for the
possibility that the optimal design lay at CM values higher than 0.85, a further seven
hulls were generated with CM values varying between 0.845 and 0.875. Three of
these hulls are shown in Figure 88. It can be seen that the bilges of these boats are
firm, but not to an extreme degree.
Figure 88. High CM designs based on parent model 1
These seven hulls were analysed using SPLASH, their results combined with the
previous training set, and neural network training performed. Only two
optimisation runs could be performed in the small amount of time remaining prior
to commencing work on parent model 2, with the results listed as runs 11 and 12 in
Table 20. Although the recommended BWL for run 11 was the same as for run 9, the
recommended CM for both runs increased to 0.87.
· 171 ·
Wallying tests
As the option for wallying in variable wind direction is included in VESPA’s race
model, a brief investigation was performed into the effect of wallying on the
optimal design parameters.
One optimisation run was performed with wallying turned on, run number 12 in
Table 20. Wind velocity parameters were the same as for run 11. However, the wind
direction oscillated with a SD of 7 degrees. Although only one run was performed
due to time constraints, it resulted in a small increase of 11mm in the optimal BWL.
This result was interesting but not conclusive, and other tests involving wallying
were scheduled for later optimisation runs.
Conclusions – parent model 1
VESPA located two optimal designs having lower BWL and higher CM values than
parent model 1. These optimised designs were successfully validated using SPLASH
analysis to provide comparisons with the SPLASH results of their parent models
The results of this testing raised the possibility that, in the absence of trailing boat
penalties, the solution space is bi-modal i.e. having two distinct optimal designs.
This situation may occur due to the format of the racecourse used, which has equal
amounts of upwind and downwind sailing.
Design features that result in a fast boat upwind usually conflict with the factors
that result in a fast boat downwind. In an event where an equal amount of time is
spent on upwind and downwind legs, this may result in a situation where a boat
that is fast upwind but slow downwind have similar elapsed time around the course
when compared to a boat that is slow upwind but fast downwind.
This appears to be the case with designs R1 and R2. Despite having similar
performance for an upwind/downwind course, these boats show different biases for
upwind and downwind work. R1 has greater beam with zero flare and does well
upwind while performing acceptably downwind. R2 is a narrower boat with 7.8
degrees of flare, and does very well downwind yet comparatively poorly upwind.
Factors that appear to favour one optimum over the other are the wind velocity
variance and the degree of benefit conferred on the boat that leads around the first
windward mark. The bimodality disappeared when an upwind bias in the form of a
trailing boat penalty was applied, resulting in a single optimal design, which
performed comparatively better upwind than down.
While flare appears to have some benefit in conditions that are highly variable, the
expected variance in the wind velocity in Valencia during the summer months did
not appear to be sufficient to justify the addition of flare to the design.
· 172 ·
9.3 parent model 2, July 30, 2006

Analysis of parent model 2 commenced in early August 2006. Parameter value
ranges for this series are shown in Table 21. The choice of CM range for parent
model 2 was made before any optimisation could be performed using the extended
parent model 1 dataset, and as a result the CM range chosen did not extend higher
than 0.87.
BWL Flare
CM CP LCB
(m) (deg.)
Minimum 2.9 0 0.80 0.58 0.525
Maximum 3.4 8 0.87 0.60 0.535
For these runs, CP and LCB were fixed to values in the centre of the ranges analysed
by SPLASH. The results of these runs, shown in Table 22, showed that, contrary to
expectations, the optimal BWL increased as mean wind velocity became stronger. It
also became clear that VESPA preferred models with CM values at the upper limit of
the parameter range available.

(m) (deg.) penalty
Run 1 9 4 1.4 2.900 0.0 0.863 0 no
Run 2 10 4 1.4 3.079 0.0 0.87 0 no
Run 3 11 4 1.4 3.112 0.0 0.87 0 no
Run 4 9 4 1.4 3.038 0.0 0.87 6 no
Run 5 11 4 1.4 3.125 0.0 0.87 6 no
Run 6 13 4 1.4 3.148 0.0 0.87 6 no
As the initial runs for parent model 2 showed little or no variation in optimal values
for either CM or flare, and as parent model 3 was due to be available for testing soon
afterwards, the opportunity to perform a sensitivity analysis for the BWL parameter
was taken. This was achieved using VESPA as a tournament-modelling program,
rather than as an optimiser, simply by limiting the number of generations to one.
For this analysis, a twenty-eight boat population was created which had identical
design parameters aside from BWL, which varied between 2.9 m and 3.4 m. CM was
fixed at 0.87 and flare set at zero degrees. Mean event wind velocity was set to 11
knots, event SD to 4 knots and race SD to 1.4 knots.
This population of yachts was analysed by VESPA, which calculated average race
times as well win/loss ratios for each boat competing in a multi-race tournament
· 173 ·
against every other boat in the population. Additionally, each boat had its win/loss
ratio calculated for a tournament against single opponents, termed exemplars. Two
different exemplars were used, one with a BWL of 3.05m and one with a BWL of 3.2m.
Results of this analysis are displayed in Figure 89, showing win/loss ratios and
average boat speed, normalised to the maximum for that optimisation run, for a
range of BWL values close to the optimal values.
versus all opponents versus BWL 3.05
versus BWL 3.2 boatspeed

Normalised win/loss ratio and boat speed
1.000
0.990
0.980
0.970
0.960
0.950
3.05 3.1 3.15 3.2
BWL
Figure 89. BWL sensitivity analysis
The optimal BWL value based on boat speed alone was 3.111m. However, when the
win/loss ratio for each competitive scenario was considered, the optimal BWL
increased to between 3.135m and 3.146m.
A second sensitivity analysis was subsequently performed for CM, which was varied
between 0.82 and 0.89. BWL was fixed at 3.05m. Results of this sensitivity analysis
are shown in Figure 90. The win/loss ratio, determined using a tournament of
match races against the entire population, is correlated with increasing CM and
reaches a maximum at 0.87, the highest CM value available for this series of designs.
· 174 ·
0.6
0.55
0.5
Win/loss ratio
0.45
0.4
0.35
0.3
0.82 0.83 0.84 0.85 0.86 0.87
CM
Figure 90. CM sensitivity analysis
Based on these results, two test hulls were generated for comparison with the parent
model using the parameters shown in Table 23. For comparison purposes, a narrow
BWL model with flare (X1) was generated to investigate the performance
characteristics of this style of hull, as there was concern that this may have been a
viable design option.
Table 23. Test Hull Parameter Values – Parent Model 2
BWL Flare
CM
(m) (deg.)
X1 2.9 0.87 7
X3 3.100 0.87 0
Results of SPLASH testing of these two models are shown in Figure 91. Model X1
showed an improvement in performance relative to the parent model in wind
velocities less than 12 knots, but performed poorly upwind in stronger winds.
Model X3, on the other hand, showed superior performance to its parent model in
all conditions upwind and in all but the strongest conditions downwind.
The performance of X3 was a pleasing confirmation that VESPA was successfully

locating designs that were superior to those arrived at by manual design methods,
within the accuracy constraints of the CFD code used. This is no guarantee that the
design would actually prove superior at full scale, in actual sailing conditions, but
locating a superior design within the limits of the design tools available was
nonetheless a significant step forward.
· 175 ·
Upwind
1
0.5
-0.5
-1
-1.5 Parent Model 2

X1
-2
X3
-2.5
6 8 10 12 14 16 18 20
Wind speed (knots)
Downwind
0.75
0.25
-0.25
-0.75
-1.25
-1.75 Parent Model 2

X1
-2.25
X3
-2.75
6 8 10 12 14 16 18 20
Wind speed (knots)
1
0.5
-0.5
-1
-1.5 Parent Model 2

X1
-2
X3
-2.5
6 8 10 12 14 16 18 20
Wind speed (knots)
Figure 91. Comparison plots of X1 and X3 versus parent model 2
· 176 ·
9.4 parent model 3, August 18, 2006

The final parent model, known within Alinghi as MS, was optimised during a 9 day
period in August 2006. Thirty-two design variations (25 training models and 7 test
models) based on the parent were generated, and subsequently analysed using
SPLASH. Parameter ranges for the training models are shown in Table 24.
BWL Flare
CM CP LCB
(m) (deg.)
Minimum 2.9 0 0.80 0.575 0.520
Maximum 3.4 4 0.89 0.595 0.535
Neural network metamodels were created using the SPLASH output and
optimisation runs performed for a variety of conditions using VESPA. Results for
these runs are shown in Table 25. Rather than locking CP and LCB to specific values
for these runs, a small range for each variable was allowed with upper limits of CP =
0.592 and LCB= 0.532
Mean Trailing
Event Race
wind BWL Flare boat
SD SD CM CP LCB
velocity (m) (deg.) penalty
(knts) (knts) (knts) (s)
Run 1 9 4 1.4 3.026 0.862 0.589 0.532 0 6
Run 2 10 4 1.4 3.042 0.863 0.589 0.532 0 6
Run 3 11 4 1.4 3.068 0.867 0.589 0.531 0 6
Run 4 13 4 1.4 3.110 0.877 0.589 0.532 0 6
Run 5 15 4 1.4 3.175 0.859 0.590 0.529 0 6
An evaluation of the optimal design using an exemplar, based on a probable

competitor for the America’s Cup final, was performed. Additionally, an evaluation
was performed against a range of exemplars, which were based on a range of likely
designs from challenging teams.
Table 26. Parent Model 3 Tests Versus Exemplars, Mean Wind Velocity = 11 knots,
Event SD = 4 knots and Race SD = 1.4 knots
BWL Flare
CM CP LCB
(m) (deg.)
Versus exemplar
BWL 3.05 3.075 0 0.88738 0.592 0.532
CM 0.885
Versus all 3.09 0 0.88877 0.592 0.532

opponents
· 177 ·
These two experiments resulted in small increases in both BWL and CM, consistent
with the hypothesis that VESPA would seek to gain a small advantage upwind over
any specific exemplars.
An evaluation of Wallying was also performed for this parent design across a range
of wind strengths to determine whether a significant effect was apparent. To
simplify the analysis these tests were performed using fixed values for some design
variables, with flare = 0°, CP = 0.59 and LCB = 0.532. Tests were performed with 100%
wallying to exaggerate any differences. Test results are summarised in Table 27.
Table 27. Effects of Wallying on design optima – Parent Model 3
Trailing
Mean wind Event Race Wind BWL boat
velocity SD SD direction SD CM
(m) penalty
(knts) (knts) (knts) (degr.)
(s)
Run 6 9 4 1.4 10 3.025 0.868 6
Run 7 11 4 1.4 10 3.075 0.871 6
Run 8 13 4 1.4 10 3.119 0.867 6
These tests show that Wallying results in a small increase, approximately 5-10 mm,
in the optimal BWL. This difference is not sufficient to warrant the consideration of
wallying as a factor in the design of an ACC yacht. This is particularly true for
VESPA, as accounting for wallying requires the calculation, for a given wind velocity,
of a portion of the yacht’s performance polar-curve over a range of angles, rather
than the calculation of a single VMG value. This significantly reduces the
performance of VESPA with little benefit in return.
9.5 Analysis of results
9.5.1 Sensitivity of design factors
The following are parameter sweeps produced by VESPA, keeping all parameters
constant except for one free parameter, which was varied through a range of values.
Baseline design parameters were BWL = 3.1m, CM = 0.87, flare = 0°, CP = 0.589 and
LCB = 0.532. These sweeps were generated using a mean wind velocity of 11 knots,
an event SD of 4 knots and a race SD of 1.4 knots.
Values on the Y axis are seconds per mile of VMG relative to the baseline model, with
lower values better than higher values. Other than the win/loss ratio curve shown in
Figure 94, the vertical scales for all graphs show seconds per mile speed penalty
relative to the optimal design. Vertical scales are equalised in order to simplify
comparisons between the effects of different parameter variations.
· 178 ·
Note that results are averages, taken from a stochastic model having a large number
of samples. As a result, the curves produced are not perfectly smooth as might be
expected if a deterministic model was used. Figure 92 displays the effect of variation
in CM on performance. This indicates a clear preference for higher CM values,
although there appears to be negligible benefit above a CM of 0.87.
Figure 92. CM sensitivity analysis
Figure 93 displays the effect of flare variation, showing a disadvantage for anything
but the zero flare state. One anomaly apparent in this graph is the dip in the curve
in the region of 5 to 6 degrees of flare. This may be interpreted as a remnant of the
bimodality previously described. It is likely that if the wind velocity variance used
for sensitivity analysis had been higher, the dip in the curve may have extended
further towards the abscissa.
Figure 93. Flare sensitivity analysis
The final sweep, shown in Figure 94, is for variation in BWL. This shows a preference
for a beam, based on raw times, of approximately 3.025 m. However, when win/loss
ratio is examined, the optimal beam increases to approximately 3.070 m.
· 179 ·
Figure 94. BWL sensitivity analysis
Note that while the win/loss ratio curve appears to be similar to an inverted curve of
raw times, there is a subtle difference in the shape of the curves. The peak of the
win/loss curve is shifted to the right by approximately 50mm, as well as having a
greater decline in value once a BWL of 3.19 is exceeded.
9.5.2 Choice of parameters for tank testing
The result of this analysis was that two optimised designs were delivered to the
Alinghi design team as candidates for a tank test model. The design parameters for
these two boats are listed in Table 28.
Table 28. Tank Model Candidate Parameter Values – Parent Model 3
BWL Flare
CM CP LCB
(m) (deg.)
Model A 3.100 0 0.87 0.589 0.532
Model B 3.140 0 0.87 0.589 0.532
While VESPA indicated that the narrower boat, Model A, was the faster of the two,
Model B was considered a more versatile design in the event that wind conditions
were significantly different from those expected during the final of the America’s
Cup. Model B was on the right shoulder of the optimal BWL curve and was
considered a more robust solution, particularly upwind. As a result, model B was
selected as the basis for the construction a 1/3 scale tank-test model, shown in
Figure 95.
· 180 ·
Figure 95. Tank model MT
9.5.3 Validation against tank test data
Tank testing of both the optimised model (MT) and parent model (MS) was
performed during September 2006. Results are displayed in Figure 96 as
comparison plots. These show performance comparisons calculated using SPLASH,
as well as performance based on tank test data. The comparison baseline for each
set of results is data from the equivalent source for the MS model, shown as a
horizontal line.
The key feature of these results is not the specific performance of the MT model,
which showed itself to be faster upwind in all wind velocities and slower downwind
in light weather than MS. Rather, it is the close correspondence of the performance
differences for the tank results and the SPLASH results that are significant. There is
good agreement between SPLASH and tank data for the speed differences that MT
would exhibit over its parent model.
· 181 ·
Upwind
0.5
Advantage/disadvantage (m/minute) MS
MT SPLASH
0
MT Tank
-0.5
-1
-1.5
6 8 10 12 14 16 18 20
Wind speed (knots)
Downwind
1.5
MS
MT SPLASH
1
MT Tank
0.5
-0.5
6 8 10 12 14 16 18 20
Wind speed (knots)
1
MS
MT SPLASH
0.5
MT Tank
-0.5
-1
6 8 10 12 14 16 18 20
Wind speed (knots)
Figure 96. MS parent model versus MT, SPLASH and tank results comparison
· 182 ·
9.5.4 Outcome
The MT tank model showed significant improvement over its parent model, MS.
Unfortunately, when the MS model was tank tested, it was not found to be faster
than the previous best design, MO. As a result, MT also did not show itself to be
superior to MO.
This left the Alinghi design team in a difficult position. The recomendations made
by VESPA across three parent models were remarkably consistent, with optimal
values for each being close to 3.1m BWL and 0.87 CM. These parameters, known
within the Alinghi design team as the X3 transformation, after the first design
produced by VESPA that showed a significant advantage over its parent design, were
considered by many to be applicable to any of the Alinghi design candidates.
As a result, a design was generated taking MO as the parent model and applying the
X3 transformation. The resulting design, MOX3, was analysed in SPLASH and PAP
with encouraging results (Figure 97). Upwind performance in all conditions was
excellent, while MO was slightly faster downwind in most conditions. Aggregate
performance upwind and downwind was superior to MO in all wind strengths.
Examination of the performance comparison plots for designs optimised by VESPA

using parent models 1 and 2 indicate that the optima found were biased towards
lighter winds than their parents. This corresponded with the opinion of the Alinghi
designers, who recognised that the performance profile of their second boat, SUI-
100, should be focussed on lighter conditions than their first boat, SUI-91.
For parent model 3, this situation was reversed, with VESPA improving the
performance in moderate to strong winds at the cost of some reduction in
performance in lighter winds. This indicated that parent model 3 had been pushed
too far in the direction of light wind performance. It is significant that the upwind
performance curve of MOX3 is almost parallel to that of MO, indicating that
VESPA concurred with the Alinghi designers as to the ideal performance profile for
the final design.
Conversely, VESPA clearly preferred to trade off downwind performance in return

for upwind performance for MOX3 relative to MO. Whether such a trade off
would have had a positive influence on the outcome of the America’s Cup final
remains a matter for conjecture.
Regardless of any potential advantage shown by SPLASH testing, MOX3 had not
been tank tested, and no time was available for further tank testing or design
development. No matter how promising the design may have been based on CFD
testing, it was universally agreed that the risks inherent in building a boat without
confirmation from tank testing were unacceptable. As a result, MO was chosen as
the design for SUI-100, the boat that subsequently won the 2007 America’s Cup.
· 183 ·
Upwind
0.5
MO
-0.5
MO X3
-1
-1.5
6 8 10 12 14 16 18 20
Wind speed (knots)
Downwind
1
0.5
-0.5
MO
MO X3
-1
6 8 10 12 14 16 18 20
Wind speed (knots)
0.5
-0.5
-1
MO
MO X3
-1.5
6 8 10 12 14 16 18 20
Wind speed (knots)
Figure 97. MO versus transformed design MOX3
· 184 ·
9.6 Summary
VESPA was used to make design recommendations based on three different parent
designs. During this period, complications resulting from the competitive fitness
function used by VESPA were identified, and steps taken to limit their adverse
effects. These issues included:
• Recognition of the implications of game theory in relation to the choice of
yacht design parameters for match racing. This suggests that the interactions
between boats when approaching rounding marks results in a bias in favour of
the leading boat, and that this bias may prevent a Nash equilibrium occurring
for match races between specific yacht designs.
• Implementation of both a simple Hall-of-Fame method and optional use of
exemplar designs by the GA.
The importance of high-quality metamodels was also recognised, with procedures

put in place to ensure that metamodels had sufficiently low prediction error while
avoiding problems with over-fitting.
Various investigations into factors that may have affected the choice of design
parameters were performed, including:
• evaluation of the effects of different wind distributions, involving changes to
both the mean wind velocity and its standard deviation, on the optimal design
configuration;
• investigation into the effect of trailing boat penalties on the optimal design
configuration;
• evaluation of the effect of wallying on the optimal design configuration.
Designs considered optimal by VESPA were re-analysed using SPLASH, and in the
case of parent model 3, by tank testing of a model based on the optimised hull
shape. Results from these analyses were compared to equivalent data derived from
the parent models to verify that VESPA had been able to make genuine design
improvements.
Despite significant variations in parent models used, the results provided by VESPA
were remarkably consistent. The design parameter values recommended for each
parent model are similar, and where differences do occur they are primarily because
of variations in the design characteristics of the parent model.
· 185 ·
Chapter 10 · Conclusions
10  Conclusions
The VESPA project was intended to produce an automated optimisation system that
could locate the optimal design parameters for an ACC yacht for a given set of
weather conditions, while executing in a reasonable time on commonly available
personal computer hardware.
In order to achieve these objectives, many disparate components had to be

integrated into a functional and reliable system. These components comprised:
• Sampling method. A modern Design of Experiments method, the Uniform
Design, was selected from a range of candidate DOE methods, based on the
high degree of uniformity exhibited by its sampling.
• Measure of merit. Rather than restrict the measure of merit to what was
achievable using a particular optimisation approach, the aim for VESPA was to
determine the ideal measure of merit and to design the system around it. The
choices made in the design of VESPA are a consequence of the characteristics
of this measure of merit,
Rather than using a measure of merit based on the performance of a single
boat, or on the results of a race between two boats, VESPA implements a
tournament model where a boat’s fitness is based on its performance against
multiple boats, across multiple races, in a range of weather conditions.
• ACC rule constraints. The optimisation process undertaken by VESPA had
the luxury of being unconstrained and having a single objective, due both to
· 186 ·
the nature of the measure of merit used and the form of the ACC rule. The
ACC rule acts as a set of penalty constraints on the design variables, ensuring
that only rule-feasible combinations of design variables are evaluated. All
designs evaluated by VESPA are checked by a code module that verifies and, in
conjunction with the parametric transformation function, enforces ACC rule
compliance.
• Parametric transformation. The parametric transformation function was
required to satisfy a broad range of requirements. Transformed hulls must
match the sometimes-conflicting set of design parameter values, while
retaining the fairness and convexity characteristics of the parent hull, and
simultaneously satisfying the constraints of the ACC rule. To satisfy these
demands, an innovative parametric transformation method was devised. This
combined powerful capabilities for the deformation of hull geometry with a
concise numerical representation, ideally suited for use by an optimisation
algorithm.
• CFD Analysis. A proven potential-flow CFD program, SPLASH, was used to
provide hydrodynamic analysis of the various hull designs provided by the
parametric transformation function.
• Metamodel selection and fitting. A review of recent research into the use of
data approximation methods revealed that many candidates for metamodel
representations exist. The determination of which method is best suited to a
particular application must be based on a variety of factors, not least of which
are the validation and verification features of the software used to create the
metamodels. In the case of VESPA, an evaluation of the requirements revealed
that artificial neural networks were best suited to the specific problem by
virtue of their approximation properties, their ease of use and their availability
in powerful commercially available software packages.
• Metamodel re-creation. To allow the neural-network metamodels to be used
within VESPA, as well as by a variety of spreadsheet macros used for testing
and validation of the system, a software module was written that could re-
create a particular neural network from a definition contained in an XML file
• Performance prediction. VESPA was able to utilise an established
performance prediction program supplied by the Alinghi team. This program,
PAP, was modified to allow it to use neural-network metamodels as a source of
hydrodynamic data, and to allow it to be called from the VESPA race-
modelling program as a dynamically linked library.
• Race modelling. The VESPA race-modelling program was designed to be an
accurate yet efficient representation of the simplest race that would correctly
rank a pair of yachts. This was a departure from the trend toward increasingly
detailed simulations of physical systems, necessitated by the requirement for
computational efficiency, as well as by the pitfalls of attempting to model
· 187 ·
highly complex systems in detail. While it is not feasible to accurately quantify

the random variation in each of the multiple components of a detailed
simulation, it is possible to make a reasonable estimate of the overall variance,
and this approach has been adopted for the VESPA race model.
• Tournament modelling. VESPA uses a measure of merit based on a
tournament of races featuring multiple boats sailing in multiple races, across a
range of weather conditions. This approach provides a superior estimate of the
probability of a yacht winning a series of races, particularly when the
population can be seeded with individuals based on the design parameters of
known or expected competitors.
• Optimisation algorithm. A review of optimisation methods indicated that a
genetic algorithm was best suited to the problem posed by VESPA, due to the
ability of the GA to handle multi-modal, non-linear solutions, and the need to
optimise a population of candidate designs.
• Testing and validation of components. Each component created for VESPA
was extensively tested for correct functioning, and validated to ensure accuracy
of results
• Testing and validation of results. No simulation of a complex, real-world
system can be one-hundred percent accurate, and for the optimisation of non-
linear systems, a small inaccuracy can result in a large error. All design
recommendations produced by VESPA were checked using SPLASH and PAP
to verify their performance. Importantly, this verification was performed by
members of the Alinghi design team in Spain, making it an independent test
of the system’s efficacy.
Once assembled and tested, VESPA worked remarkably well, showing itself capable
of making genuine improvements to highly developed parent models, using modest
computer resources. Optimised designs were shown to be superior to their parent
designs when directly re-analysed using SPLASH, and in one case, when tank tested
at one-third scale. Execution time for an optimisation run was reasonable, taking
approximately 12 hours on a 2.4 GHz Pentium based system.
10.1 Critical assessment & contribution

The development of VESPA required the integration of many complex components.
Failure of any one of these components to perform as required would have rendered
the system unworkable. Significant verification and validation of each step in the
process was implemented to ensure a viable optimisation system. This degree of
rigour was considered essential, as the goal of the research was a system capable of
producing design recommendations of value to the designers of the America’s Cup
defender.
· 188 ·
This research endeavoured to achieve objectives that had not previously been
attempted. Although other researchers had devised yacht-design optimisation
systems that solved simplified subsets of the problem, the scope of work involved in
the creation of VESPA was substantial, and the goals of the system ambitious.
The design of ACC yachts has progressed over a period of 15 years, to the point
where current designs are highly optimised and the scope for further improvement
in is small. Consequently, the accuracy required of VESPA’s analysis and simulation
tools was high, as the smallest of errors might result in erroneous outcomes from the
optimisation process.
Despite undertaking a complex real-world optimisation problem, involving designs

that had already undergone thousands of hours of design development, numerical
analysis and tank testing, VESPA was able to make a meaningful contribution to the
design development process of the Alinghi team. While limited to some extent by
the accuracy of the CFD tools available, VESPA was able to achieve verifiable
improvements to the highly refined parent models supplied by Alinghi.
The research described by this thesis makes significant contributions to the field of
racing-yacht design and to the field of naval architecture in general. The use of
metamodels for complex non-linear hydrodynamic data has been shown to be a
viable approach, as has the sparse, quasi-random sampling of the design space to
provide a concise basis for those metamodels. A novel parametric transformation
method has been presented that combines a parsimonious set of control parameters
with a powerful set of geometric controls, permitting variation of hull volumes
longitudinally, transversely and vertically without violating fairness and convexity
conditions. Lastly, the combination of an efficient race simulation with a
tournament-modelling algorithm permitted the evaluation of a racing yacht in the
most appropriate manner - via a comparison with a diverse assembly of its peers.
10.2 Advantages and disadvantages of VESPA

Currently the weak link in the methodology is the accuracy of the results produced
by the CFD analysis, as small errors in this area can result in misleading outcomes
from the optimisation process. However, CFD is a field undergoing constant
development and it is likely that accuracy will improve in the future.
Alinghi also has access to two RANS codes, FLUENT and CFX. It was considered
that the execution times of these programs were too long given the number of data
points needing to be calculated, while the improvement in accuracy would be small.
Although these programs can be run on a cluster of computers, the performance
differential compared to SPLASH running on a single CPU is still significant.
· 189 ·
It is possible, due to improvements in both hardware and software, that RANS codes
may be sufficiently fast for use with VESPA at some time in the near future. If this
was to occur, the change could be easily accommodated within VESPA, with the
data from the RANS code being indistinguishable from data derived from SPLASH.
Little or no change to the data preparation sequence would be required.
The use of metamodels by VESPA is an approach whose value increases as execution

times for the CFD analysis grow larger. Consequently, it is anticipated that the use
of metamodels will be of even greater benefit when used with RANS codes whose
execution times are substantially higher than potential flow codes such as SPLASH.
Although VESPA has been specifically targeted at ACC design, it is equally

appropriate to other forms of yacht design, including long distance ocean racing
and fleet racing. In these cases, the objective function may need to be changed from
being tournament based to being based on aggregate course times, and a different
optimisation method may be more appropriate than the GA currently used by
VESPA. However, use of non-linear metamodels to approximate the results of an
expensive CFD analysis, while using a stochastic RMP to model the tournament
structure of the actual racing, would remain as principal features of the system.
10.3 Future work and outlook

Although the scope of work undertaken in the creation of the VESPA system was
broad, not all aspects of the problem could be addressed without raising the level of
complexity above what was practicable within the time and resources available. As a
result, there are many refinements that may be made to VESPA in the future. These
include:
• adaptation of the method to vessels other than ACC yachts;
• incorporation of potentially superior CFD methods, such as Reynolds averaged
Navier-Stokes (RANS) codes;
• incorporation of other metamodelling techniques such as radial basis function
networks (RBF);
• investigation of automated, variable-fidelity models;
• parallelisation of the genetic algorithm to allow it to run on multiple CPU
cores simultaneously;
• blending of multiple parents as a complementary design variation method;
• extension of VESPA to encompass other design aspects, such as sail and rig
optimisation, or multidisciplinary problems such as keel, bulb and rudder
design. Currently the VESPA system is focussed on hull shape optimisation
· 190 ·
only, however there is no reason why other components of a racing yacht

should not be part of the optimisation process;
• incorporation of additional hydrodynamic analysis methods, such as added
resistance in waves, into the simulation model. These refinements would, in
turn, require the variation of additional design parameters, such as the
independent movement of LCF relative to the LCB location. Additional
hydrodynamic effects such as added resistance could be satisfied using direct
analysis methods, or alternatively, using additional pre-calculated metamodels.
The America’s Cup is currently in hiatus, due to legal challenges that have resulted
in a competition between only two teams, to be sailed in large multihulls during
February 2010. Once this event is concluded, it is likely that the America’s Cup will
return to ballasted monohulls, but in a new class that will replace the current
America’s Cup Class yachts.
A new class will provide a significant opportunity for the further development and
use of VESPA. Rather than coaxing minute improvements from designs that are
already close to optimal due to being governed by a mature rating rule, all designers
will be working from a clean sheet of paper. In this situation, VESPA may be able to
dramatically shorten design development time by rapidly isolating the most
promising regions of the design space and exploiting them in an expeditious manner.
· 191 ·
Publications
Publications
The following are publications by the author, relevant to the subject matter
contained in this thesis:
Couser, P., Mason, A. P., Von Konsky, B., Smith, C. & Mason, G. 2004,
'Artificial neural networks for hull resistance prediction', COMPIT 04,
Siguenza, Spain.
Mason, A. P., Couser, P., Von Konsky, B., Smith, C. & Mason, G. 2005,
'Optimisation of vessel resistance using genetic algorithms and artificial
neural networks', COMPIT 05, Hamburg, Germany.
Mason, A. P. & Thomas, G. 2007, 'Stochastic optimisation of IACC yachts',

COMPIT 07, Cortona, Italy.
Mason, A. P. 2008, 'Stochastic optimisation for America’s Cup Class yachts based
on a genetic algorithm', Schiffstechnik, vol. 55, pp. 60-77.
· 192 ·
References
References
Abt, C., Bade, S. D., Birk, L. & Harries, S. 2001a, 'Parametric hull form design - a
step towards one week ship design', PRADS 2001, Shanghai.
Abt, C., Harries, S., Heimann, J. & Winter, H. 2003, 'From redesign to optimal
hull lines by means of parametric modeling', COMPIT 03, Hamburg.
Abt, C., Harries, S. & Hochkirch, K. 2001b, 'Efficient geometric modeling in

modern yacht design', Symposium für Yachtbau und Yachtkonstruktion,
Hamburg.
Abt, C., Harries, S. & Hochkirch, K. 2004, 'Constraint management for marine
design applications', 9th Symposium on Practical Design of Ships and Other
Floating Structures, Luebeck-Travemuende, Germany.
ACC 2003, America's Cup Class rule version 5.0.
Akaike, H. 1974, 'A new look at the statistical model identification', IEEE
Transactions Auto. Control., vol. 19, pp. 716-723.
Alkan, A. D., Gulez, K. & Yilmaz, H. 2004, 'Design of a robust neural network
structure for determining initial stability particulars of fishing vessels', Ocean
Engineering, vol. 31, no. 5-6, pp. 761-777.
Alonso, J. J., LeGresley, P. & Pereyra, V. 2009, 'Aircraft design optimization',

Applied Mathematics and Computers in Simulation, vol. 79, no. 6, pp. 1948-
1958.
Alyuda 2005, Neurointelligence Manual.
Baik, J. & Gonella, P. 2005, Automatic optimisation of IACC yacht hulls, M.Sc. thesis,
Chalmers University of Technology, Göteborg, Sweden.
Bales, N. K. 1980, 'Optimizing the seakeeping performance of destroyer-type hulls',

13th ONR Symposium on Naval Hydrodynamics, Tokyo, Japan.
Bezier, P. 1974, 'Mathematical and practical possibilities of UNISURF', in Computer

Aided Geometric Design, eds. R. E. Barnhill & R. F. Riesenfeld, Academic
Press, New York.
Birk, L. 2005, 'Parametric modelling and optimization of offshore structure', ICCAS

2005, Busan.
· 193 ·
References
Birk, L. & Harries, S. 2000, 'Automated optimization - a complementing technique

for the hydrodynamic design of ships and offshore structures', COMPIT 2000,
Potsdam/Berlin, Germany.
Birmingham, R. & Smith, T. A. G. 1998, 'Automatic hull form generation: a

practical tool for design and research', Practical Design of Ships and Mobile
Units, The Hague.
Birmingham, R., Webster, B., Roskilly, T. & Jones, E. 2002, 'The application of
artificial intelligence to roll stabilisation for a range of loading and operating
conditions', HISWA 2002, The International Symposium on Yacht Design and
Yacht Construction.
Bole, M. & Lee, B. S. 2006, 'Integrating parametric hull generation into early stage
design', Schiffstechnik, vol. 53, pp. 115-137.
Box, G. E. P. & Wilson, K. B. 1951, 'On the experimental attainment of optimum

conditions (with discussion)', Journal of the Royal Statistical Society, Series B
(Methodological) vol. 13, no. 1, pp. 1-45.
Burges, C. J. C. 1998, 'A tutorial on support vector machines for pattern

recognition', Data Mining and Knowledge Discovery, vol. 2, no. 2, pp. 121-
167.
Campana, E. F., Fasano, G., Peri, D. & Pinto, A. 2007, 'Nonlinear programming
approaches in the multidisciplinary design optimization of a sailing yacht keel
fin', 9th International Conference on Numerical Ship Hydrodynamics, Ann
Arbor, Michigan.
Campana, E. F., Peri, D., Tahara, Y. & Stern, F. 2006, 'Shape optimization in ship
hydrodynamics using computational fluid dynamics', Journal of Computer
Methods in Applied Mechanics and Engineering, vol. 196, no. 1-3, pp. 634-651.
Černý, V. 1985, 'A thermodynamical approach to the travelling salesman problem:

an efficient simulation algorithm', Journal of Optimization Theory and
Applications, vol. 45, no. 1, pp. 41-51.
Chance, B., Jr. 1987, 'The design and performance of Twelve Meter yachts',
Proceedings, American Philosophical Society, vol. 131 no. 4
Chen, B., Zhao, C. & Qiu, X. 1996, 'Global optimization of ship main dimensions
by simulated annealing algorithm', Shipbuilding of China, vol. 2, no. 133, pp.
16-26.
Chen, E. & Parent, R. 1989, 'Shape averaging and its applications to industrial
design ', IEEE Computer Graphics and Applications vol. 9, no. 1, pp. 47-54.
· 194 ·
References
Chen, P.-F. 2004, The inverse design studies in estimating the optimal hull form, Ph.D.
thesis, National Cheng Kung University, Tainan, Taiwan.
Chen, V. C. P., Tsui, K.-L., Barton, R. R. & Allen, J. K. 2003, 'A review of design
and modeling in computer experiments', Handbook of Statistics, no. 22, pp.
231-261
Clarey, C. 2000, 'Stars and Stripes loses on land and on the sea ', NY Times, January
10, 2000.
Claughton, A. R. & Oliver, J. C. 2003, 'Development in hydrodynamic force

models for velocity prediction programs', R.I.N.A. International Conference on
the Modern Yacht.
Couser, P., Mason, A. P., Von Konsky, B., Smith, C. & Mason, G. 2004, 'Artificial
neural networks for hull resistance prediction', COMPIT 04, Siguenza, Spain.
Cramer, N. L. 1985, 'A representation for the adaptive generation of simple

sequential programs', International Conference on Genetic Algorithms and Their
Applications, Pittsburgh, PA.
Day, A. H. 1993, 'Steps towards an optimal sailplan', RINA Spring Meeting 1993.
Day, A. H. & Doctors, L. J. 1997, 'Resistance optimization of displacement vessels

on the basis of principal parameters', Journal of Ship Research, vol. 41, no. 4,
pp. 249-259.
Day, A. H. & Doctors, L. J. 2000, 'The survival of the fittest - evolutionary tools
for hydrodynamic design of ship hull form', RINA Transactions,
De Boor, C. & Ron, A. 1990, 'On multivariate polynomial interpolation',

Constructive Approximation, vol. 6, no. 3, pp. 287-302.
de Jong, E. D. 2004, 'Intransitivity in coevolution', Parallel Problem Solving from

Nature - PPSN VIII, Birmingham, UK.
de Jong, K. A. 1975, An analysis of a behaviour of a class of genetic adaptive systems,

Ph.D. thesis, University of Michigan, Ann Arbor.
DeBord, F., Reichel, J., Rosen, B. & Fassardi, C. 2002, 'Design optimization for the
International America’s Cup Class', SNAME Transactions, vol. 110
Dejhalla, R., Mrša, Z. & Vukovic, S. 2001, 'Application of genetic algorithm for
ship hull form optimization', International Shipbuilding Progress, vol. 48, no.
2, pp. 117-133.
· 195 ·
References
Delhommeau, G. 1987, Les problèmes de diffraction-radiation et de résistance de

vagues: Etude théorique et résolution numérique par la méthode des singularités,
Thèse de doctorat d'état, Université de Nantes, Nantes.
Desideri, J.-A., Duvigneau, R., Abou El Majd , B. & Tang, Z. 2006, 'Algorithms for
efficient shape optimization in aerodynamics and coupled disciplines', 42nd
AAAF Congress on Applied Aerodynamics, Sophia-Antipolis, France.
Duvigneau, R. 2007, Aerodynamic shape optimization with uncertain operating

conditions using metamodels, No.6143, INRIA.
Duvigneau, R. & Visonneau, M. 2001, 'Towards a practical design optimization

tool for incompressible and turbulent flows', 4th Numerical Towing Tank
Symposium, Hamburg, Germany.
Duvigneau, R. & Visonneau, M. 2002, 'Hybrid genetic algorithms and neural

networks for fast CFD-based design'.
Duvigneau, R., Visonneau, M. & Deng, G. B. 2002, 'On the role played by
turbulence closures in hull shape optimization at model and full scale', 24th.
ONR Symposium on Naval Hydrodynamics.
Dyn, N., Levin, D. & Rippa, S. 1986, 'Numerical procedures for surface fitting of
scattered data by radial basis functions', SIAM Journal of Scientific and
Statistical Computing, vol. 7, no. 2, pp. 639-659.
Eberhart, R. C. & Kennedy, J. 1995, 'A new optimizer using particle swarm theory',
Sixth International Symposium on Micromachine and Human Science, Nagoya,
Japan.
Eiben, A. E. & Schippers, C. A. 1998, 'On evolutionary exploration and

exploitation', Fundamenta Informaticae, vol. 35
Evans, J. H. 1959, 'Basic design concepts', Naval Engineers Journal, vol. 71, no. 4,
pp. 671-678.
Fang, K.-T. 2004, Uniform experimental design tables,

http://www.math.hkbu.edu.hk/UniformDesign/main.html#ud_tables,
accessed September 2005.
Fang, K.-T. 2006, 'Recent development in the uniform experimental design', Int.
Conf. on Design of Experiments and its Applications, Tianjin
Fang, K.-T., Lin, D. K. J., Winker, P. & Zhang, Y. 2000, 'Uniform design: theory
and application', Technometrics, vol. 42, no. 3 pp. 237-248.
· 196 ·
References
Fassardi, C. & Hochkirch, K. 2006, 'Sailboat design by response surface

optimization', 2nd High Performance Yacht Design Conference, Auckland, New
Zealand.
Fisher, R. A. 1935, The design of experiments, Oliver and Boyd, Edinburgh.
Fletcher, R. 1987, Practical methods of optimisation, Wiley.
Fogel, L. J., Owens, A. J. & Walsh, M. J. 1966, Artificial intelligence through

simulated evolution, Wiley.
Formation Design Systems 2006, Maxsurf User Manual, version 12, Formation
Design Systems.
Friedman, J. H. 1991, 'Multivariate adaptive regression splines (with discussion)',

Annals of Statistics, vol. 19, no. 1, pp. 1-141.
Galan, M., Winter, G., Montero, G., Greiner, D., Periaux, J. & Mantel, B. 1996, 'A
transonic flow shape optimisation using genetic algorithms', ECCOMAS
conference on numerical methods in engineering No.2, Paris, France.
Gerritsma, J., Onnink, R. & Versluis, A. 1981, 'Geometry, resistance and stability
of the Delft yacht hull series', International Shipbuilding Progress, vol. 28, pp.
276-297.
Giunta, A. A. & Wojtkiewicz, S. F. J. 2003, Overview of modern design of experiments

methods for computational simulations, AIAA 2003-0649, Sandia Laboratories.
Goldberg, D. & Richardson, J. 1987, 'Genetic algorithms with sharing for

multimodal function optimization', 2nd International Conference on Genetic
Algorithms.
Goldberg, D. E. 1989, Genetic Algorithms in search, optimization and machine

learning, Addison-Wesley, 412 pages.
Greenman, R. M. 1998, Two-dimensional high-lift aerodynamic optimization using

neural networks, NASA/TM-1998-112233, NASA Ames Research Center,
Moffett Field, California.
Grefenstette, J. J. 1986, 'Optimization of control parameters for genetic algorithms',

IEEE Transactions on Systems, Man and Cybernetics, vol. 16, no. 1, pp. 122-
128.
Gretzky, J. A. & Marshall, J. K. 1993, 'The Partnership for America's Cup

Technology: an overview', 11th Chesapeake Sailing Symposium, Annapolis.
· 197 ·
References
Halley, D. 1987, On the systematic variation of hull representations for computers,

Royal Institution of Naval architects.
Harries, S. & Abt, C. 1999a, 'Formal hydrodynamic optimization of a fast

monohull on the basis of parametric hull design', 5th International Conference
on Fast Sea Transportation (FAST 1999), Seattle.
Harries, S. & Abt, C. 1999b, 'Parametric design and optimization of sailing yachts',
14th Chesapeake Sailing Symposium, Annapolis.
Harries, S., Abt, C., Heimann, J. & Hochkirch, K. 2006, 'Advanced hydrodynamic
design of container carriers for improved transport efficiency', Design &
Operation of Container Ships, London.
Harries, S., Abt, C. & Hochkirch, K. 2001a, 'Hydrodynamic modeling of sailing

yachts', 15th Chesapeake Sailing Yacht Symposium, Annapolis.
Harries, S., Birk, L. & Abt, C. 2003a, 'Parametric hull design - the FRIENDSHIP-
modeler', NAV 2003. Palermo, pp.
Harries, S., Janson, C. E., Leer-Anderson, M., Marzi, J., Maisonneuve, J. J., Raven,
H. & Valdenazzi, F. 2003b, 'The FANTASTIC Ro-Ro: CFD optimisation of
the forebody and experimental verification', NAV 2003, Palermo.
Harries, S., Valdenazzi, F., Abt, C. & Viviani, U. 2001b, 'Investigation on

optimization strategies for the hydrodynamic design of fast ferries', 6th
International Conference on Fast Sea Transportation (FAST 2001),
Southampton.
Hearn, G. E. & Wright, P. N. H. 1997, 'Seakeeping for design: optimisation of

motion responses and wave making resistance of catamarans via the
application of a genetic algorithm', 4th International Conference on Fast Sea
Transportation (FAST 1997), Sydney.
Heimann, J. & Harries, S. 2003, 'Optimization of the wave-making characteristics

of fast ferries', 7th International Conference on Fast Sea Transportation (FAST
2003), Ischia, Italy.
Hendrix, D., Percival, S. & Noblesse, F. 2001, 'Practical hydrodynamic

optimization of a monohull', SNAME Transactions, vol. 109, pp. 173-183.
Hess, D. E., Faller, W. E., Ammen, E. S. & Fu, T. C. 2004, 'Neural networks for
naval applications', COMPIT 04, Siguenza, Spain.
Hess, J. L. & Smith, A. M. O. 1966, 'Calculation of potential flow about arbitrary

bodies', Progress in Aeronautical Science, pp. 1-138.
· 198 ·
References
Hestenes, M. R. & Stiefel, E. 1952, 'Methods of conjugate gradients for solving

linear systems', Journal of Research of the National Bureau of Standards, vol. 49,
no. 6, pp. 409-436.
Hirata, N. 2004, 'Comparison of genetic algorithm and gradient-based method to

ship shape optimization', Journal of Kansai Society of Naval Architects, no. 241,
pp. 59-65.
Hochkirch, K., Röder, K., Abt, C. & Harries, S. 2002, 'Advanced parametric yacht
design', High Performance Yacht Design Conference, Auckland.
Holland, J. 1975, Adaptation in natural and artificial systems, University of

Michigan Press, Ann Arbor.
Holtrop, J. & Mennen, G. G. J. 1978, 'A statistical power prediction method',

International Shipbuilding Progress, vol. 25 no. 290, pp. 253-256.
Hooke, R. & Jeeves, T. A. 1961, '"Direct search'' solution of numerical and

statistical problems', Journal of the ACM, vol. 8, no. 2, pp. 212-229.
Hsiung, C.-C. 1981, 'Optimal ship forms for minimum wave resistance', Journal of
Ship Research, vol. 25, no. 2, pp. 95-116.
Hsiung, C.-C. & Shenyan, D. 1984, 'Optimal ship forms for minimum total
resistance', Journal of Ship Research, vol. 28, no. 3, pp. 163-172.
Hsiung, C. & Shenyan, D. 1985, 'Applying resistance theory and quadratic

programming method to determine optimal ship forms', Workshop on
Developments in Hull Form Design, Wageningen, Netherlands.
Hurrion, R. D. & Birgil, S. 1999, 'A comparison of factorial and random

experimental design methods for the development of regression and neural
network simulation metamodels', The Journal of the Operational Research
Society, vol. 50 no. 10 pp. 1018-1033.
IMS 2004, IMS - International Measurement System, a handicapping system for

cruising/racing yachts.
Islam, M. M., Khondoker, M. R. H. & Rahman, C. M. 2001, 'Application of

artificial intelligence techniques in automatic hull form generation', Ocean
Engineering, vol. 28, no. 12, pp. 1531-1544.
Jacquin, E., Alessandrini, B., Bellevre, D. & Cordier, S. 2002, 'Yacht optimisation
based on genetic algorithm and RANSE solver', High Performance Yacht
Design Conf., Auckland.
· 199 ·
References
Jagoda, J. 1973, 'Computer-aided multi-level optimization method applied to

economic ship design', International Conference on Computer Applications in
the Automation of Shipyard Operation and Ship Design, Tokyo, Japan.
Janikow, C. Z. & Michalewicz, Z. 1991, 'An experimental comparison of binary

and floating point representations in genetic algorithms', Fourth International
Conference on Genetic Algorithms.
Janson, C. & Larrson, L. 1996, 'A method for the optimization of ship hulls from a
resistance point of view', 21st Symposium on Naval Hydrodynamics,
Trondheim, Norway.
Jin, R., Chen, W. & Simpson, T. W. 2003, 'Comparative studies of metamodeling

techniques under multiple modeling criteria', AIAA 2000.
Jin, Y., Olhofer, M. & Sendhoff, B. 2001, 'Managing approximate models in

evolutionary aerodynamic design optimization', IEEE Congress on
Evolutionary Computation, Seoul, South Korea.
Jin, Y., Olhofer, M. & Sendhoff, B. 2002, 'A framework for evolutionary
optimization with approximate fitness functions', IEEE Transactions on
Evolutionary Computation, vol. 6, no. 5, pp. 481-494.
Johnson, M., Moore, L. & Ylvisaker, D. 1990, 'Minimax and maximin distance
design', Journal of Statistical Planning and Inference, vol. 26, pp. 131-148.
Keane, A. J. 1996, A brief comparison of some evolutionary optimization methods,

Department of Engineering Science, University of Oxford, Oxford.
Keane, A. J. & Nair, P. 2005, 'Elements of numerical optimization', in

Computational Approaches for Aerospace Design: The Pursuit of Excellence, John
Wiley & Sons, p. 602.
Keane, A. J., Price, W. G. & Schachter, R. D. 1991, 'Optimisation techniques in

ship concept design', RINA Transactions, Pt A, vol. 133, pp. 123-144.
Kerwin, J. E. 1978, A velocity prediction program for ocean racing yachts, revised to
June 1978, Report No.78-11, Massachusetts Institute of Technology.
Keuning, J. A., Gerritsma, J. & Van Terwigsa, P. F. 1993, 'Resistance test of a series
of planing hull forms with a 30 degree deadrise angle, and a calculation
model based on this and similar systematic series', International Shipbuilding
Progress, vol. 40, no. 424, pp. 333-385.
· 200 ·
References
Keuning, J. A. & Sonnenberg, U. B. 1999, 'Approximation of the calm water

resistance on a sailing yacht based on the Delft systematic yacht hull series',
14th Chesapeake Sailing Yacht Symposium, Annapolis.
Kirkpatrick, S., Gelatt, C. D. & Vecchi, M. P. 1983, 'Optimization by simulated

annealing', Science, vol. 220, no. 4598, pp. 671-680.
Korpus, R. 2004, 'Reynolds-Averaged Navier-Stokes in an integrated design

environment.', MDY 04.
Korpus, R. 2007, 'Performance prediction without empiricism: a RANS-based VPP

and design optimization capability', 18th Chesapeake Sailing Yacht Symposium,
Annapolis.
Koushan, K. 2003, Artificial neural networks for prediction of ship resistance and
wetted surface area, Norway.
Koza, J. R. 1992, Genetic programming: on the programming of computers by means of

natural selection, MIT Press, Cambridge, MA.
Kupras, L. K. 1976, 'Optimization method and parametric study in precontracted

ship design', International Shipbuilding Progress, vol. 23, pp. 138-155.
Lackenby, H. 1950, 'On the systematic geometrical variation of ship forms',

Transactions of The Institute of Naval Architects, vol. 92, pp. 289-315.
Lengyel, A. 2003, 'Ship design and product modelling using the NAPA System',
SCHIP en WERF de ZEE, pp. 22-24.
Letcher, J. S. 1974, 'Handicapping rules and performance of sailing yachts', 1st

Chesapeake Sailing Yacht Symposium, Annapolis, MA.
Letcher, J. S., Marshall, J. K., Oliver, J. C. & Salvesen, N. 1987, 'Stars & Stripes',
Scientific American, vol. 257, no. 2, pp. 34-40.
Lloyd, B. 1995, 'New Zealand races away from Conner for 3-0 lead in the Cup', NY
Times, May 10, 1995.
Maisonneuve, J. J. 1989, Résolution du problème de la résistance de vagues des navires

par une méthode de singularité de Rankine, Ph.D thesis, Ecole nationale
supérieure de mécanique de l'université de Nantes, Nantes.
Maisonneuve, J. J. 1993, 'Optimization tools for ship resistance and seakeeping

problems', Fast '93, 2nd Intl Conf on Fast Sea Transportation, Yokohama,
Japan.
· 201 ·
References
Maisonneuve, J. J. 2003, 'Application examples from industry: America's Cup hull

shape optimization', Optimistic, Berlin.
Markov, N. E. & Suzuki, K. 2001, 'Hull form optimization by shift and

deformation of ship sections', Journal of Ship Research, vol. 45, no. 3, pp.
197-204.
Mason, A. P., Couser, P., Von Konsky, B., Smith, C. & Mason, G. 2005,
'Optimisation of vessel resistance using genetic algorithms and artificial
neural networks', COMPIT 05, Hamburg, Germany.
McKay, M. D., Conover, W. J. & Beckman, R. J. 1979, 'A comparison of three

methods for selecting values of input variables in the analysis of output from
a computer code', Technometrics vol. 21, no. 2, pp. 239-245.
McNaull, R. 1980, 'Generating new ship lines from a parent hull', REAPS Technical
Symposium, National Shipbuilding Research Program, Philadelphia,
Pennsylvania.
Menzel, S., Olhofer, M. & Sendhoff, B. 2005, 'Application of free form

deformation techniques in evolutionary design optimisation', 6th World
Congress on Structural and Multidisciplinary Optimization, Rio de Janeiro.
Mesbahi, E. & Atlar, M. 2000, 'Artificial neural networks: applications in marine

design and modelling', COMPIT 2000, Potsdam.
Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H. & Teller, E.

1953, 'Equation of state calculations by fast computing machines', Journal of
Chemical Physics vol. 21, no. 6, pp. 1087-1092.
Michell, J. H. 1898, 'The wave resistance of a ship', Philosophical Magazine of the

Royal Society Ser.5, vol. 45, pp. 106-123.
Milgram, J. H. 1993, 'Naval architecture technology used in winning the 1992

America's Cup match', SNAME Transactions, vol. 101, pp. 399-436.
Minami, Y. & Hinatsu, M., , 2002, 'Multi-objective optimization of ship hull form
design by response surface methodology', 24th Symposium on Naval
Hydrodynamics, Fukuoka, Japan.
Moore, G. E. 1965, 'Cramming more components onto integrated circuits',

Electronics, vol. 38, no. 8
Morgan, T. H. 1916, A critique of the theory of evolution, Princeton University Press.
Morishita, M. & Akagi, S. 2005, 'Hull form optimization of fast ships using
simulated annealing (SA) method', COMPIT 05, Hamburg.
· 202 ·
References
Morris, T. D. & Mitchell, T. J. 1995, 'Exploratory designs for computational

experiments', Journal of Statistical Planning and Inference, vol. 43, no. 3, pp.
381-402.
Muhlenbein, H. & Schlierkamp-Voosen, D. 1993, 'Predictive models for the

breeder genetic algorithm', Evolutionary Computation, vol. 1, no. 1
Nash, J. 1950, Non-cooperative Games, Ph.D. thesis, Princeton University.
Nelder, J. A. & Mead, R. 1965, 'A simplex algorithm for function minimization',
Computer Journal,
Neocleous, C. C. & Schizas, C. N. 1995, 'Artificial neural networks in marine

propeller design', ICNN, 1995
Neu, W. L., Good, N. & Klasen, E. 2007, 'The effect of synthesis model
uncertainties in ship design optimization', COMPIT 07, Cortona, Italy.
Neu, W. L., Hughes, O., Mason, W. H., Ni, S., Chen, Y., Ganesan, V., Lin, Z. &
Tumma, S. 2000a, 'A prototype tool for multidisciplinary design
optimization of ships', Ninth Congress of the International Maritime
Association of the Mediterranean, Naples, Italy.
Neu, W. L., Mason, W. H., Ni, S., Lin, Z., Dasgupta, A. & Chen, Y. 2000b, 'A
multidisciplinary design optimisation scheme for container ships', 8th
AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis and
Optimisation, Long Beach, CA.
Niederreiter, H. 1992, 'Monte Carlo and quasi-Monte Carlo methods in scientific

computing', CBMS-NSF, vol. 63
Noblesse, F. 1983, 'A slender-ship theory of wave resistance', Journal of Ship

Research, vol. 23, no. 1, pp. 13-33.
Notz, W. 2003, 'Experimental designs for computer experiments', Experimental

Design Workshop, Academia Sinica, Taipei, Taiwan.
Nowacki, H., Brusis, F. & Swift, P. M. 1970, 'Tanker preliminary design-an

optimization problem with constraints', SNAME Annual Meeting.
Nowacki, H., Creutz, G. & Munchmeyer, F. 1977, 'Ship lines creation by computer
- objectives, methods and results', First International Symposium on Computer-
Aided Hull Surface Definition (SCAHD77), Annapolis, MD.
Nowacki, H. & Kim, H. C. 2005, 'Form parameter based design of hull shapes as
volume and surface object', ICCAS 2005, Busan, Korea.
· 203 ·
References
Nowacki, H. A. 1983, 'Design and fairing of ship surfaces', in Surfaces in Computer

Aided Geometric Design, ed. R. a. B. W. Eds. Barnhill, North Holland, pp.
121-134.
Nowacki, H. A. 1993, 'Hull form variation and evaluation', Journal of the Kansai
Society of Naval Architects, no. 219 pp. 173-184
Ockam 2006, OckamU manual, Ockam Instruments.
Oei, C. K., Goldberg, D. E. & Chang, S. J. 1991, Tournament selection, niching,

and the preservation of diversity, 91011, University of Illinois at Urbana-
Champaign, Urbana, Illinois.
Oliver, J. C. & Claughton, A. R. 1995, 'Development of multi-functional velocity

prediction program for sailing', Int. Conf. on Computer Aided Design and
Production for Small Craft, CADAP '95, Southampton.
Oliver, J. C., Letcher, J. S. & Salvesen, N. 1987, 'Performance prediction for Stars
& Stripes', SNAME Transactions, vol. 95, pp. 239-261.
Ong, Y. S., Nair, P. B. & Keane, A. J. 2003, 'Evolutionary optimization of

computationally expensive problems via surrogate modeling', AIAA Journal,
vol. 41, no. 4, pp. 687-696.
Oortmerssen, G. 1971, 'A power prediction method and its application to small
ships', International Shipbuilding Progress, vol. 18, no. 207
Panait, L. & Liuke, S. 2002, 'A comparison of two competitive fitness functions',
GECCO 2002.
Park, J.-S. 1994, 'Optimal Latin-hypercube designs for computer experiments',

Journal of Statistical Planning and Inference, vol. 39, no. 1, pp. 95-111.
Parolini, N. & Quarteroni, A. 2007, Modelling and numerical simulation for yacht
engineering, MOX-Report No.10/2007, Department of Mathematics,
Politechnico di Milano, Milan.
Parsons, M. G. 1975, 'Optimization methods for use in computer-aided ship

design', First Ship Technology and Research (STAR) Symposium, Washington
D.C.
Peacock, D., Smith, W. F. & Pal, P. K. 1996, Hull-form generation using multi-
objective optimisation techniques, AME CRC C 96/11, Australian Maritime
Engineering CRC.
Percival, S., Hendrix, D. & Noblesse, F. 2001, 'Hydrodynamic optimization of ship

hull forms', Applied Ocean Research, vol. 23, no. 6, pp. 337-355.
· 204 ·
References
Peri, D. & Campana, E. F. 2005a, 'Global optimization for safety and comfort',
COMPIT 05, Hamburg.
Peri, D. & Campana, E. F. 2005b, 'High-fidelity models and multiobjective global

optimisation algorithms in simulation-based design', Journal of Ship Research,
vol. 49, no. 3, pp. 159-175.
Peri, D. & Campana, E. F. 2005c, 'High fidelity models in global optimization',

COCOS 2003, Berlin / Heidelberg.
Peri, D. & Campana, E. F. 2005d, 'High fidelity models in simulation-based design',
Peri, D., Campana, E. F. & Di Mascio, A. 2001, 'Development of CFD-based

design optimization architecture', First MIT Conference on Computational
Fluid and Solid Mechanics.
Peri, D. & Mandolesi, F. 2005, 'Multiobjective design optimization of an IACC

sailing yacht by means of CFD high fidelity solvers', 17th Chesapeake Sailing
Yacht Symposium, Annapolis.
Philpott, A. B. 2003, 'Stochastic optimization in yacht racing', in Applications of

Stochastic Programming, eds. W. Ziemba & S. Wallace, Springer-Verlag.
Philpott, A. B., Henderson, S. G. & Teirney, D. P. 2004, 'A simulation model for
predicting yacht match race outcomes', Operations Research, vol. 52, no. 1, pp.
1-16.
Piegl, L. & Tiller, W. 1997, The NURBS book, Springer-Verlag, New York.
Pierret, S. 1999, 'Turbomachinery blade design using a Navier-Stokes solver and

artificial neural network', ASME Journal of Turbomachinery, vol. 121, no. 3,
pp. 326-332.
Pinto, A. & Campana, E. F. 2005, 'A multi-swarm algorithm for multi-objective

ship design problems', COMPIT 05, Hamburg.
Pinto, A., Peri, D. & Campana, E. F. 2004, 'Global optimization algorithms in

naval hydrodynamics', COMPIT 04, Siguenza, Spain.
Pinto, A., Peri, D. & Campana, E. F. 2007, 'Multiobjective optimization of a

containership using deterministic particle swarm optimization', Journal of
Ship Research, vol. 51, no. 3, pp. 217-228.
Poloni, C., Giurgevich, A., Onesti, L. & Pediroda, V. 1999, 'Hybridisation of a

multi-objective genetic algorithm, a neural network and a classical optimizer
for a complex design problem in fluid dynamics',
· 205 ·
References
Poloni, C. & Pediroda, V. 1997, 'GA coupled with computationally expensive

simulations: tools to improve efficiency', in Genetic algorithms in Engineering
and Computer Science, eds. J. Periaux, D. Quagliarella, C. Poloni & G.
Winter, John Wiley & Sons.
Powell, M. J. D. 1987, 'Radial basis functions for multivariable interpolation: a

review', in Algorithms for Approximation, eds. J. C. Mason & M. G. Cox,
Oxford University Press, London.
Quagliarella, D. & DellaCioppa, A. 1994, 'Genetic algorithms applied to the

aerodynamic design of transonic airfoils', 12th AIAA Applied Aerodynamics
Conference, Colorado Springs, CO.
Queipo, N., Haftka, R., Shyy, W., Goel, T., Vaidyanathan, R. & Tucker, K. 2005,
'Surrogate-based analysis and optimization', Progress in Aerospace Sciences, vol.
41, no. 1, pp. 1-28.
Rechenberg, I. 1973, Evolutionsstrategie: optimierung technischer systeme nach

prinzipien der biologischen evolution, Frommann Holzberg Verlag, Stuttgart.
Reenskaug, T. M. H. 2003, 'Applications and technologies for maritime and

offshore industries', History of Nordic Computing Conference, Trondheim.
Reichel, J., Pugh, J., Rosen, B. S. & Debord, F. 1994, 'Grand prix yacht design with
the aid of computational and experimental techniques', Yacht Vision '94
Symposium, Auckland, New Zealand.
Rogers, D., F. 1977, 'B-spline curves and surfaces for ship hull design', First
International Symposium on Computer-Aided Hull Surface Definition
(SCAHD77), Annapolis, Maryland.
Rosen, B. S., Laiosa, J. P. & Davis, W. 2000, 'CFD studies for America's Cup 2000',
AIAA-2000-4339,
Rosen, B. S., Laiosa, J. P., Davis, W. H., Jr. & Stavetski, D. 1993, 'SPLASH free-
surface flow code methodology for hydrodynamic design and analysis of
IACC yachts', 11th Chesapeake Sailing Yacht Symposium, Annapolis.
Rosin, C. & Belew, R. 1996, 'New methods for competitive coevolution',

Evolutionary Computation, vol. 5, no. 1, pp. 1-29.
Rudolph, G. 1994, 'Convergence analysis of canonical genetic algorithms', IEEE

Transactions on Neural Networks, vol. 5, pp. 96-101.
· 206 ·
References
Rumelhart, D. E., Hinton, G. E. & Williams, R. J. 1986, 'Learning internal

representations by error propagation', in Parallel Distributed Processing:
Explorations in the Microstructures of Cognition, MIT Press, Cambridge, MA,
pp. 318-362.
Sacks, J., Welch, W. J., Mitchell, T. J. & Wynn, H. P. 1989, 'Design and analysis of
computer experiments ', Statistical Science, vol. 4, no. 4, pp. 409-423.
Sahoo, P. 1997, 'Determination of hull form by use of non-linear optimization

methods', IV International Symposium on High-Speed Marine Vehicles, Naples,
Italy.
Salvesen, N., Tuck, E. O. & Faltinsen, O. 1970, 'Ship motions and sea loads', Trans.
SNAME, vol. 78, pp. 250-287.
Sánchez, P. J. 2006, 'As simple as possible, but no simpler: A gentle introduction to

simulation modeling', Winter Simulation Conference 2006, Monterey,
California.
Sarioz, K., Hearn, G. E. & Hills, B. 1992, 'Practical seakeeping for design: an
optimised approach', PRADS '92, 5th Intl. Symp. on the Practical Design of
Ships and Mobile Units, Newcastle upon Tyne, U.K.
Sarle, W. S. 1994, 'Neural networks and statistical models', Nineteenth Annual SAS
Users Group International Conference, Cary, NC.
Savitsky, D. 1964, 'Hydrodynamic design of planing hulls', Marine Technology,
Schaffer, J. D., Caruana, R. A., Eshelman, L. J. & Das, R. 1989, 'A study of control
parameters affecting online performance of genetic algorithms for function
optimization', Third International Conference on Genetic Algorithms, George
Mason University, United States.
Schlageter, E. C. & Teeters, J. R. 1993, 'Performance prediction software for IACC

yachts', 11th Chesapeake Sailing Yacht Symposium, Annapolis, MD.
Schwefel, H.-P. 1974, Adaptive mechanismen in der biologischen evolution und ihr
einfluß auf die evolutionsgeschwindigkeit, Abschluflbericht zum DFG-Vorhaben
Re 215/2, Technical University of Berlin.
Sederberg, T. W. & Parry, S. R. 1986, 'Free-form deformation of solid geometric

models', ACM Siggraph Computer Graphics, vol. 20, no. 4, pp. 151-160.
Seif, M. S. & Jahanbakhsh, E. 2004, 'Neural networks model for ship maneouver',
COMPIT 04, Siguenza, Spain.
· 207 ·
References
Sen, P. & Todd, D. S. 1997, 'Multiple criteria genetic algorithms: a catamaran

design study', 7th International Conference on Genetic Algorithms, East Lansing,
MI.
Simpson, T. W., Korte, J. J., Mauery, T. M. & Mistree, F. 1998, Comparison of

response surface and kriging models for multidisciplinary design optimization,
4755, AIAA.
Simpson, T. W., Lin, D. K. J. & Chen, W. 2000, 'Sampling strategies for computer
experiments: design and analysis', International Journal of Reliability and
Applications,
Simpson, T. W., Peplinski, J. D., Koch, P. N. & Allen, J. K. 2001, 'Metamodels for
computer-based engineering design: survey and recommendations', in
Engineering with Computers, Springer-Verlag, pp. 129-150
Sobol, I. M. 1967, 'On the distribution of points in a cube and the approximate
evaluation of integrals', U.S.S.R. Computational Mathematics and Mathematics
Physics, vol. 7, pp. 86-112.
Söding, H. & Rabien, U. 1977, 'Hull surface design by modifying an existing hull',
First International Symposium on Computer-Aided Hull Surface Definition
(SCAHD77), Annapolis, MD.
Tan, C., Ray, T. & Tsai, H. 2003, 'A comparative study of evolutionary algorithm
and swarm algorithm for airfoil shape optimization problems', 41st Aerospace
Sciences Meeting and Exhibit, Reno, Nevada.
Tang, B. 1993, 'Orthogonal array based Latin hypercube sampling', Journal of

American Statistical Association, pp. 1392-1397.
Teeters, J. 2004, IMS performance package, U.S. Sailing - Offshore Racing Council.
Teirney, D. P. 1998, 'Yacht match race simulation', 33rd Annual Conference of the
ORSNZ, University of Canterbury
Thompson, D. 1917, On Growth and Form, Cambridge University Press.
Valdenazzi, F., Harries, S., Viviani, U. & Abt, C. 2002, 'Seakeeping optimisation of
fast vessels by means of parametric modelling', HSMV 02, Napoli, Italy.
Valdenazzi, F., Pittaluga, C., Harries, S., Abt, C. & Avellino, G. 2003,
'Hydrodynamic design of the aftbody shape of a roro vessel', NAV 2003,
Palermo.
van Oossanen, P. 1985, 'The development of the 12 Meter class yacht “Australia II”',
7th Chesapeake Sailing Yacht Symposium.
· 208 ·
References
van Oossanen, P. 1993, 'Predicting the speed of sailing yachts', SNAME Transactions,
vol. 101, pp. 337-397.
Volker, H. 1954, 'Systematic variation of usual ship form', International

Shipbuilding Progress, vol. 1, no. 1
Walden, D. A., Kopp, P. J. & Grundmann, P. 1985, 'Optimization of hull form for
seakeeping and resistance', Workshop on Developments in Hull Form Design,
Wageningen, Netherlands.
Whitfield, R. I., Hills, B. & Coates, G. 1999, 'The application of multi-objective

robust design methods in ship design', 10th International Conference on
Computer Applications in Ship Building, (ICCAS'99), Cambridge, USA.
Whitfield, R. I., Wright, P. N. H., Coates, G. & Hills, W. 1998, 'A robust design
methodology suitable for application to one-off products', Journal of
Engineering Design, vol. 9, no. 4, pp. 373-387.
Wright, A. H. 1991, 'Genetic algorithms for real parameter optimization', in

Foundations of Genetic Algorithms, Morgan Kaufmann, pp. 205-218.
Yamamoto, K. & Inoue, O. 1995a, 'Applications of genetic algorithm to

aerodynamic shape optimization', 12th AIAA Computational Fluid Dynamics
Conference, San Diego.
Yamamoto, K. & Inoue, O. 1995b, 'New evolutionary direction operator for

genetic algorithms', AIAA Journal, vol. 33, no. 10, pp. 1990-1992.
Yang, C., Noblesse, F. & Lohner, R. 2001, 'Practical hydrodynamic optimization of

a trimaran', SNAME Annual Meeting, Orlando, FL.
Yang, C., Noblesse, F., Lohner, R. & Hendrix, D. 2000, 'Practical CFD applications
to design of a wave cancellation multihull ship', 23rd Symposium on Naval
Hydrodynamics, Val De Reuil, France.
Yasukawa, H. 2000, 'Ship form improvement using genetic algorithm', Ship

Technology Research, vol. 47, no. 1, pp. 35-44.
Zhou, Z., Ong, Y. S., Nair, P. B., Keane, A. J. & Lum, K. Y. 2007, 'Combining
global and local surrogate models to accelerate evolutionary optimization',
IEEE Transactions on Systems, Man and Cybernetics, Part C: Applications and
Reviews, vol. 37, no. 1, pp. 66-76.
Zitzler, E. 2002, 'Evolutionary algorithms for multiobjective optimization',

Evolutionary Methods for Design, Optimisation and Control, Barcelona, Spain
· 209 ·

Americas Cup Optimisation

Uploaded by

Copyright:

Available Formats

Americas Cup Optimisation

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Americas Cup Optimisation

Uploaded by

Copyright:

Available Formats

S TOCHASTIC OPTIMISATION OF

A MERICA’ S C UP C LASS YACHTS

Andrew Phillip Mason

A THESIS SUBMITTED IN FULFILMENT OF THE REQUIREMENTS FOR

AUSTRALIAN MARITIME COLLEGE

_________________________ February 12, 2010

Statement of authority of access

VESPA proved itself capable of making genuine design improvements to an existing

Thanks also to Barbara Backhouse for her meticulous proofreading skills.

In short, a vessel must be light enough to be driven easily by a moderate breeze,

The satisfaction of any one requirement necessitates something antagonistic to

4.5.4 Parametric transformation 71

8 Verification & Validation 139

Figure 63. Deformed FFD lattice P’ with updated weights ′, resulting in

Nomenclature – naval architecture

Frd depth Froude number, /√

FB distance from the forward perpendicular to centre of buoyancy

LCF longitudinal centre of flotation

Nomenclature – yacht design

βTW true wind angle

AAE Average Absolute Error (∑ | |) /

ARE Average Relative Error (∑ | |

Acronyms – design & analysis

Acronyms - optimisation methods

Acronyms - approximation methods

America’s Cup Class - rule variables

ABC aft beam correction

Prior to the introduction of Computational Fluid Dynamics (CFD) software, yacht

If a workable design optimisation system could be developed, several benefits would

Once a design is close to optimal, it becomes more difficult to discover design

In addition, if the system is able to perform global optimisation, thoroughly

1.1 Problem definition

The first step in such an optimisation problem is determining how to ascertain

1.2 Background and context

Racing yacht design adds further complexity, with a mix of performance

Figure 1. Design Spiral

Within the field of optimisation, this approach is referred to as the alternating

An example of the alternating variable method incorporating two parameters is

Figure 2. Alternating variable method

This leads to oscillatory behaviour of the alternating variable algorithm, illustrated

Figure 3. Alternating variable method with correlated variables

The alternating variable method is seldom used as a computer based optimisation

In contrast, many computer based optimisation algorithms, for example the

1.3 Optimisation – a definition

An optimisation problem with a single objective can be stated as:

having the following components:

Measure of merit or objective function: criteria by which a solution is assessed:

Constants: values that do not change during the optimisation process.

1.4 Design optimisation – state of the art

1.4.1 Early ship design optimisation

Table 1. Early Optimisation Studies in Basic Ship Design

Authors Year Ship Type Criterion Approach

Parsons (1975) was among the first to publish a comprehensive review of

1.4.2 Analysis methods

Computational Fluid Dynamics - panel methods

Advanced CFD methods in optimisation

An example of the performance penalties incurred by modern RANS codes is

Clearly, a global optimisation method requiring hundreds or thousands of RANS

Figure 4. SHIPFLOW zones, from Janson and Larsson (1996)

Campana et al. (2006) described several innovations designed to improve the