BOILER ERECTION SCHEDULING USING PRODUCT MODELS AND

CASE-BASED REASONING

Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

By Ren-Jye Dzeng 1 and Iris D. Tommelein,z Associate Member, ASCE

ABSTRACT: Contractors who repeatedly construct facilities designed by copying major parts from one project
to the next find that previously developed schedules associated with those designs could be reused to schedule
new work. To facilitate such reuse, project characteristics must be articulated and associated schedules described
to include not only traditional, numerical scheduling data, but also scheduling constraints. In addition, knowledge
about how to reuse schedules must be available. The CasePlan system, presented here, supports and augments
the scheduling activity of people who reason about cases-each case describing the design and schedule of a
completed project-to generate new project schedules. Specifically, CasePlan reuses annotated cases to automatically schedule the erection of power plant boilers. Be~ause such ~oile~s ~av~.a more-or-less .standardiz~d
design, a generic boiler product model can serve as the basis for assessmg similanties between designs. On this
basis CasePlan selects schedules for reuse. A user can also interact with CasePlan to isolate fragments of case
schedules and adapt them to better suit the variables of the new project at hand.

INTRODUCTION
Architects/engineers (AlEs) who repeatedly design facilities
of the same kind may reuse designs as a whole or in part. This
is the case, for instance, in cookie-cutter design of industrial
facilities such as power plants. When parts of a design are
copied from one project to the next, contractors who have
experience in constructing those facilities find that (parts of)
schedules can also be reused to schedule new construction
work. This saves time, which is of the essence during bid
preparation, but, more importantly perhaps, it enables them to
reuse some of their company's proven field expertise that is
reflected in those schedules.
A scheduler intent on reusing an existing schedule will try
to recall major parts from a past project's design and schedule
(including components and details, construction resource
availability, contractual agreements, environmental factors,
etc.) that recur in or in some way resemble the new project at
hand, in order to determine which (parts of) schedules could
be reused. The scheduler may then choose a single existing
schedule for adaption to the new design, contract, and construction needs, or choose parts from several existing schedules, link them, and adapt them to better suit those needs.
This problem-solving method-namely, reasoning about
past cases then retrieving and adapting them to solve a new
problem-is termed case-based reasoning (CBR). It is a
method used by schedulers who can remember salient features
about past projects and who have enough such projects to draw
upon. CBR works well when cases are organized so that the
search for and adaption of relevant cases can proceed in a
systematic way. Unfortunately, most valuable project knowledge gets buried in vast amounts of documentation of a paperbased archival system. Moreover, novice schedulers rarely
have access to relevant, historic case data because it is in the
head of their seniors and not accessible in any other way than
through verbal communication. A lot of knowledge is lost
when senior schedulers are tied up with other work, get promoted, leave the organization to take on other duties, or retire.
'Assoc. Prof., Civ. Engrg. Dept., National Chiao Tung Univ., Hsinchu,
30050 Taiwan.
2 Acting Assoc. Prof., Civ. and Envir. Engrg. Dept., 215-A McLaughlin
Hall, Univ. of California, Berkeley, Berkeley, CA 94720-1712.
Note. Discussion open until February I, 1998. To extend the closing
date one month, a written request must be filed with the ASCE Manager
of Journals. The manuscript for this paper was submitted for review and
possible publication on September 9, 1996. This paper is part of the
JourlUll of Construction Engineering and MalUlgement, Vol. 123, No.
3, September, 1997. ASCE, ISSN 0733-9364/97/0003-0338-0347/
$4.00 + $.50 per page. Paper No. 14048.

The aim of the research described in this paper was to address (I) how cases could be described in a systematic way to
facilitate classification and reuse; and (2) how a computerbased system could automate the generation of construction
schedules by reusing cases that reflect design and construction
experience.

RELATED WORK
Automated Planning Systems
Construction planning consists of defining activities with
their durations, precedence relationships, and resources,
whereas scheduling involves applying the critical-path method
(CPM) to calculate early and late activity start and finish times,
and floats. Computer tools that perform CPM calculations are
widely used but few exist that address the planning task. Plan
generation has been automated using artificial intelligence (AI)
programming techniques. Examples of such AI-based planning
systems (AI-planners, in short) are BUILDER (Cherneff et al.
1991), PLANEX (Zozaya-Gorostiza et al. 1989), GHOST (Navinchandra et al. 1988), Know-Plan (Morad and Beliveau
1991), OARPLAN (Darwiche et al. 1989), and SIPEC (Kartam and Levitt 1990). These are constructive planning systems: they always generate a plan from scratch for each new
project. Their knowledge typically comprises (1) a (usually
functional) hierarchy of components for a particular type of
project (e.g., high-rise steel construction) in which each type
of component has a construction activity associated with it
(though in principle possible, few models present alternatives);
and (2) planning rules or constraints acquired from expert
practitioners, based on rules-of-thumb, or stated in the literature. These systems use this knowledge to pick a network of
activities for each product's component based on its association in the hierarchy, and to test the preconditions of each
activity and determine its precedence relationship relative to
other activities in the plan.
Many of these AI-planners were developed to provide designers with constructibility feedback before award of the construction contract (Gray 1986). Their knowledge includes general construction principles but rarely project characteristics
and resources such as contractor equipment, material availability, or site constraints (PLANEX is an exception). In the
process of generating solution plans, these AI-planners keep
track of all possible alternatives while pruning out only those
that do not meet the constraints. Because relatively few constraints are known at design time, they typically can output
many plans that satisfy all constraints. By leaving the user
(contractor) with the largest possible set of satisfying plans,

338/ JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997

J. Constr. Eng. Manage. 1997.123:338-347.

Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

they follow a least-commitment approach, but consequently,

the plans they generate are not "field-executable": they need
to be refined with site-specific data before they can be used to
guide field work.
In contrast, people scheduling work typically investigate
only a few alternatives but in greater detail, i.e., they follow
an early-commitment approach. For instance, after investigating two or three alternative sequences or durations of activities, a person may commit to one of them, discard the others,
and develop the remainder of the schedule based on that
choice. We are not arguing here that either least commitment
or early commitment is better. This depends on the state of
information, objectives, and tools available to the problem
solver. Tommelein et al. (1991) discuss trade-offs between the
two. However, early commitment is the approach for people
to pursue because scheduling knowledge is so varied, difficult
to spell out exhaustively, and it is project- and organizationdependent. Accordingly, people tend to become good schedulers by gaining hands-on experience on a project-by-project
basis.
Some automated early-commitment AI-planners exist,
though they tend to focus on only one specific resource, e.g.,
Thabet and Beliveau (1993) sequence activities based on space
availability. CasePlan can accommodate sequencing knowledge based on anyone or several resources. Constructive AIplanners have no means of recognizing that they planned similar projects previously or that plans can be reused. CasePlan
recognizes that one can develop cases from previous projects
and schedules, and reuse this knowledge. Our work is most
similar in approach to Miresco's (1992), but CasePlan has
many more features and capabilities than Miresco describes.

Case-Based Reasoning
The knowledge base for a CBR system comprises a set of
cases plus mechanisms for retrieving cases and adapting their
solutions to suit the new problem at hand, rather than a set of
rules and facts that make up a traditional rule-based system.
CBR has played an important role in the development of AI
and expert systems (Kolodner 1993) addressing tasks such as
planning, design, explanation, diagnosis, classification, and
natural-language parsing. To our knowledge, however, there
exist no CBR systems for generating construction schedules.
A CBR system adopts three steps in solving a problem
(Dzeng 1995):
1. Identifying and retrieving useful case(s). This requires
access to cases and knowledge to assess the appropriateness for retrieval of one case over another. It is typically
assumed that the most useful case is the one most similar
to the new problem, though the notion of similarity varies; e.g., it can be based on the match between goals or,
instead, between design attributes in the new problem
versus those in an archival case. Nevertheless, matching
and ranking requires three steps: (I) determining the correspondence of matching attributes, (2) assigning an importance weight to attributes used in assessing similarity,
and (3) determining the degree of similarity between corresponding attribute values based on some similarity
measurement (e.g., semantic hierarchies, quantitative or
qualitative ranges of values, and functional roles the values play in problem solving).
2. Reusing and adapting the retrieved case(s). Solutions
from retrieved cases seldom perfectly match but need to
be adapted to solve the new problem. One may (1) reinstantiate a partial solution from an old solution, (2) adjust
some parameters, (3) transform the solution's structure
(e.g., delete old or add new parts) before applying it to
the new problem, or (4) replay the reasoning steps used

in the old case but using data and constraints in the new
problem.
3. Categorizing and storing the new problem and solution
as a case. When a new problem's solution has been validated, a new case can be created and stored for reuse.
Organizing stored cases is not mandatory, but using some
indexing scheme (e.g., articulating goals, salient attributes of the problem, or factors that caused failure during
problem solVing) may improve the efficiency and success
rate of subsequent searches.
Most CBR systems use similar concepts and approaches,
but vary in their combination thereof to suit their domain of
application. Like other systems, CasePlan uses CBR as a
means to reuse knowledge specific to individual projects
(which is lacking from existing construction planning systems)
but it is unique in that it tackles planning problems in construction and uses product models as the basis for case organization.

Product Modeling
A case-based scheduling system should not be a stand-alone
tool, but tie into existing modeling practices in design and
construction. Establishing a representation to integrate and
exchange data spanning the lifetime of a facility that is of use
to all participants involved with a common architectural/engineering!construction (AEC) project is a widely recognized
problem, exacerbated by the use of heterogeneous computer
systems. It has driven the development of an international
standardization effort that involves people in AEC practice as
well as academia, e.g., the Special Issue on Data and Product
Modeling in the Journal of Computing in Engineering (Special
1996).
Several standardization efforts were investigated for this research, namely:
1. Plant information network (PIN). The Electrical Power
Research Institute developed a comprehensive power
plant model as a guide to specifying integrated computeraided applications ("Guidelines" 1987). Twelve volumes describing this model were reviewed (unfortunately, no electronic copy could be obtained) but the PIN
model was not adopted because it appears to have been
superseded by PlantSTEP described in the following.
2. Standard for the exchange of product model data (STEP)
and general AEC reference model (GARM). The most
significant product-model standardization effort to date
has resulted in the International Standards Organization
(ISO) draft standard 10303, named STEP (ISO I994a).
STEP is an abstract model to be refined with domainspecific details in order to reflect each industry's ontology. For instance, PlantSTEP's application protocol AP277 serves the power plant industry (ISO 1994b) and is
currently under development. PlantSTEP is most relevant
to the present research, but, unfortunately, insufficient
detail had been formalized at the onset of this research
for it to be useful. However, our modeling methodology
is compatible with PlantSTEP's.
GARM, developed specifically for AEC applications,
was the next best available STEP standard, though it also
is an abstract reference model under which product-specific models must be developed. GARM entities describe
a product in terms of product definition units (PDUs),
which can be the entire product, its components and subcomponents, or relationships between PDUs.
3. Work breakdown structures (WBSs). To elicit the terminology of boiler components used in industry practice,
WBSs were obtained from boiler manufacturers Asea

JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997/339

J. Constr. Eng. Manage. 1997.123:338-347.

Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

Brown Boveri Combustion Engineering Services, Inc.

(ABB-CE) and Zurn Industries, Inc. (Zurn). WBSs help
organize work and serve scheduling and cost-engineering
functions. However, they provide no framework for industrywide AEC data exchange. First, they do not capture all data relevant to all AEC participants. Second,
many firms use only their own, in-house WBS, which
differs from competitors', and not all use it consistently.
Nonetheless, given their availability, CasePlan's boiler
product model was based on the WBS of one boiler manufacturer and made consistent with GARM's entity representation (Dzeng and Tommelein 1993).
SCHEDULING ERECTION OF POWER PLANT BOILER

the scheduling process, before construction starts, is a

key to success.
4. Boiler components require protection from inclement
weather and must be installed in a timely fashion upon
their arrival on site so they can be tested before their
warranty expires.
5. Power plant projects are capital intensive. Because of the
time value of money, it is important that the project duration be kept short. Many power plants are therefore
developed on a fast-track schedule. Components with a
long lead time must be procured early and their arrival
on site often drives the schedule of other work.
6. Contracts typically include financial incentives to finish
the project early as this means the early start of revenuemaking operation.

Industry Practice: Case Data Collection

Studying the erection of power plant boilers was well suited
for this research because:
1. Power plant construction is reasonably complex, though
the construction methods used are routine so (parts of)
schedules are likely to be reused from one project to the
next.
2. Boiler erection is possibly the most critical activity for
timely completion of construction of a power plant, so
optimizing erection schedules is important.
3. Boiler erection includes the installation of large, heavy,
custom-made mechanical equipment. This requires expensive lifting equipment, complex rigging procedures,
and highly skilled personnel. Identifying these as part of
Grayling Generating Station
Limited Partnership
CMS Generation Co.
Decker Energy International Inc.

Black & Veatch Power

Development Corp.

TBC Joint-venture
Townsend & Bottum Inc.
The Christman Co.

Therefore, contractors are more competitive in constructing

these kinds of projects when they invest a reasonable amount
of effort in scheduling their work.
Boiler erection was also chosen for practical reasons:
1. A fair number of boilers exist whose designs (general
arrangement of components) are similar. Thus, it was anticipated that project schedules would be similar and
therefore reusable.
2. The erection schedules we obtained from industry referred to approximately 20 major boiler components
(though there may be on the order of 15,000 subcomponents), and they comprised from 60 to 90 activities.
This degree of complexity suited this research.
3. Professionals in industry were interested in this work and
Genesee Power Station
Limited Partnership
CMS Generation Co.
Black & Veatch Power Development Corp.
Genesee Power Co.

BVTBC Trl-venture
Black & Veatch Power Development Corp.
Townsend & Bottum Inc.
The Christman Co.
(Architect/Engineer and
General Contractor)

Northern Boiler Inc.

.......

Notation

Boiler
Erection
Schedule
FIG. 1.

Organization of Grayling and Genesee Projects

340/ JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997

J. Constr. Eng. Manage. 1997.123:338-347.

Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

were willing to provide engineering expertise and data.

In fact, boiler erection was suggested as a subject of
study for this research by professionals from Black &
Veatch.
1\vo boiler projects were studied in detail: the Grayling Generating Station (Grayling) ("Biomass-fired" 1993) and the
Genesee Power Station (Genesee) (Desmond 1995). Both were
turn-key projects located in southeast Michigan and owned,
designed, and built by the same parties but under different
joint-venture agreements. Fig. 1 shows their owner (represented by a rectangle), AlEs (rounded rectangle), general contractor (octagon), boiler manufacturer (hexagon), and boiler
erector (oval). Their similar project organization allowed us to
concentrate on schedule differences caused by differences in
facility design, site environment, and project specifications.
Both plants are waste-wood fueled. They generate nearly
the same power (Grayling 34 MW and Genesee 35 MW). To
lower the design cost the AlE reused Grayling's design for
Genesee, so the structures of the two plants are similar (e.g.,
most bays have the same length) albeit not identical (e.g., the
foundations are different because of different soil conditions).
In addition, they would have liked to reuse Zurn's boiler design at Genesee, but the contract was awarded through an
open-bid process and a competitor, ABB-CE, had the lowest
bid. Despite some differences in manufacturer design, resulting among other things in differing requirements on hold-out
steel during construction, the boiler sizes and configurations
are similar because they were designed to meet the same,
owner-specified constraints (including fuel and plant capacity).
Grayling's boiler erection started in May 1991 and was completed in March 1992; Genesee's started in December 1994
and was completed in October 1995.
As-built boiler erection schedules were obtained from Zurn
and ABB-CE. As is common practice, the manufacturers were
responsible not only for designing the boiler and manufacturing its components, but they also were substantially involved
with its erection. Their schedules had therefore been generated
with a considerable amount of input from the boiler erector.
Both companies used the Primavera Project Planner (P3)
software (Primavera 1993). They saved routine networksparts of larger schedules-as reusable "fragnets" as they are
called in P3. P3 did not suit our research needs, however. It
is impossible in P3 to describe relationships between product
components (e.g., economizer, drums) and activities or subnetworks (unless one treats components as resources of activities). Yet, doing so is useful when searching through a collection of schedules to find reusable pieces. Also, P3's built-in
batch language to write macros is limited. Sophisticated computation must be performed through C-language routines external to P3. Yet, annotating a schedule with formulas or other
computable expressions to express scheduling knowledge
eases sch~dule adaption for reuse.
Scheduling Knowledge

Knowledge for creating a field-executable schedule is expressed by means of constraints, which fall into six categories
(Dzeng and Tommelein 1994, after Gray 1986 and Echeverry
et al. 1991):
1. Facility components and their relationships. Components
determine which construction alternatives are feasible,
e.g., the degree of prefabrication of an economizer affects the duration of construction and the installation
equipment required. Physical and structural relationships
between design elements (support, embedment, coverage, enclosure, etc.) often govern construction sequencing, e.g., drums of a top-supported boiler cannot be in-

2.

3.

4.

5.

6.

stalled until the supporting frame is in place, but may

require that some steel be held out for access.
Resources (time, money, people, equipment, material,
and space). The availability of resources determines activity durations, imposes sequencing of otherwise concurrent activities, and affects lead times, e.g., all waterwalls of a boiler may be installed simultaneously or parts
of it concurrently depending on the availability of labor,
equipment, material, and space to conduct the work
safely and without undue interference. Similarly, trade
interaction will prevent the structural contractor from
closing the roof until the boiler erector has finished hoisting equipment through it.
Procurement and mobilization of equipment and materials, and hiring of labor. The lead time for delivery of
custom-designed mechanical equipment may dictate the
earliest project completion time, e.g., a turbine generator
may take more than a year to be manufactured and delivered to site, so its installation tends to define the critical path. A scheduler will work backward from this delivery date to set the pace for erecting the power plant
building and boiler.
Site environment. The site environment (climate, access,
topography, etc.) determines the timing and nature of
construction activities, e.g., most contractors need to
clear and grub their site and build temporary facilities;
Michigan contractors may not start foundation work before early spring to avoid frozen ground, soggy access
roads, and problems associated with placing concrete in
cold weather.
Regulations and specifications. Contracts specify that
testing and inspection be performed or permits obtained
at given times. These introduce schedule delays or mark
milestones. The. Occupational Safety and Health Administration (OSHA) (Construction 1985) regulates that
"safety nets shall be installed and maintained whenever
the potential fall distance exceeds two stories or 8 m [25
ft]." The erection of steel therefore awaits installation of
the floors two stories below, though a contractor may
"install safety net" and "remove safety net" to speed
up frame erection.
Scheduling detail and format. The type of information,
level of detail, and format of a schedule is a function of
who generates the schedule, the information available at
that time, contractual arrangements, schedule use, and
the timing and frequency of schedule generation and updating, e.g., a contractor's schedule may detail activities
for work performed with its own forces and only summarize those for subcontracted work.

Many AI-planners use "Facility components and their relationships" as the primary constraints and they encode some
"Regulations and specifications" into their heuristic rules for
sequencing activities. Most plan with different "Scheduling
detail and format" in that they generate hierarchical plans, but
the level of detail is predetermined and not for the user to
change. Only PLANEX considers "Resources" and "Site environment" constraints. None consider "Procurement" constraints. In contrast, CasePlan allows practitioners to account
for all these constraints so schedules can be better custom tailored.
CASEPLAN ARCHITECTURE AND COMPONENTS
Project Modeling Knowledge

The CasePlan architecture builds on the premise that a generic product model, comprising a hierarchy of classes that
represent abstract and physical components with has-a-com-

JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997/341

J. Constr. Eng. Manage. 1997.123:338-347.

S
':I ----

ame: string
10: string
escrlptlon: text

PDU

B
Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

Product

can-be-a ,. ...

,.

,.
I ,.
ProjectSpecs.

I
I
I
I

Schedule

--

--

~,.

Site

II-~

Ar\',
,,
I
\
I
,,
\
I
\
,
Ij1~s-:Cmpo~ent
~.

,. ,. .,./
1//
I
I-",.

string
Conatructlon- =!~e:
as-components:
'(component ...)
PDU
s-component-of: component
-component-network: component-network

Boller

Drums

l-

II

\ I

I--

Economizer II--

Stoker

'-

' ...
' .... 1

WW

I-I-I--

Instantiation

Product-1

Boller
1

Project- lSpecs 1. I-

....

Site
1

l-

I has-a-component
I
I

I
I-

I-I-I--

I-

I-

....

I-

Drums
1

-.
--

FIG. 2.

Economizer
1

IlI-

....

I
Stoker
1

I-

I-

....

I-

I
WW
1-1

~I

I
WW
1-2

--

Class Inheritance

ponent links between them, can describe the type of project

of interest (here, power plant boilers). Specifically, a generic
'Case' for a boiler comprises a 'Product' and a 'Schedule'
(Fig. 2). The product (here, 'Boiler') in turn comprises components such as 'Orums', 'Economizer', and 'Waterwall'
('WW'), but also 'Project Specifications' and 'Site'.
The Project Specifications class describes constraints of the
type "Procurement, mobilization, and hiring" and "Regulations and specifications." Similarly, 'Site' describes "Site environment" and 'Product' describes "Facility components and
their relationships" constraints. 'Schedule' comprises activity
subnetworks and precedence links between them, which controls "Scheduling detail and format." "Resources" can be
assigned to activities and their productivity used to estimate
activity durations.
The 'Case' for an individual, new project is created by instantiating the 'Project' class hierarchy, so it may comprise
zero, one, or several instances of each component (e.g., 'WW
1-1' and 'WW 1-2' both instantiate 'WW'). At first, a new
'Case' will not include a 'Schedule' instance but it will be
created by CasePlan.
In terms of implementation, 'Product', 'Schedule', 'Case',
and 'Construction POD' are classes specializing the class
'POD' (see Fig. 2 and section "Related Work"). A 'Construction-POD' has attributes: (1) type, naming the class type (e.g.,
the product 'Boiler' that describes the hierarchy as a whole,
or the component 'Economizer') of the item in the product
model; (2) has-components, listing the components of the
'Construction-POU'; (3) is-component-of, referring to the object of which the 'Construction-POD' is a component; (4)
component-network, referring to the sequence of activities that
represents the process by which to construct the 'ConstructionPOD'.
Again, a number of classes specialize 'Construction-POD'.
They inherit its attributes plus add their own. 'Site' may have

attributes: location, soil type, ground-water level, etc. and

'Project Specifications' may require activities (e.g., to include
the "Hydro test" activity that cannot be directly related to any
one boiler component, but which is to be performed when the
boiler has been substantially completed) and impose constraints (e.g., to specify that some inspection activity must
occur before a certain date, i.e., that it is subject to a timing
constraint).

Schedule Modeling Knowledge

A CPM schedule's activities and sequential links are instances of the classes 'Activity' and 'Link', both of which
specialize the 'POD' (not shown in Fig. 2).

Component Network
CasePlan constructs a schedule by determining a network
of activities (termed a "component network") describing the
construction process for each component in the product model
and then combining them into a single large network. A sequential link in a component network is of type start-to-start
(SS), start-to-finish (SF), finish-to-start (FS), or finish-to-finish
(FF). Dnless marked otherwise, a single line represents the
default FS link. Links that have no arrow imply that they go
from left to right.
Fig. 3(a) shows the component network for a typical erection sequence for 'Drums'. First, the steam drum (upper drum)
is raised to a height approximately equal to the length of the
generating bank while the mud drum (lower drum) is raised
just above the ground. With the drums temporarily secured,
pipefitters assemble the generating bank, which connects the
drums, near the ground. When the bank assembly (including
both drums and therefore referred to as the component 'Orums',
in the plural) is completed, it is further raised. After the drums
have been erected, pipefitters finish their piping hookups. Fig.

342/ JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997

J. Constr. Eng. Manage. 1997.123:338-347.

(a)

Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

Raise
steam drum
(Drums)

10-

Ground assemble
Erect
Raise
mud drum f-- generating bank ,..- drums
(Drums)
(Drums)
(Drums)

1\

Trim
external piping
(Drums)

Trim
internal piping
(Drums)

Trim
external piping
(Drums)

(b)
Erect
Erect
mud drum f-- steam drum
(Drums)
(Drums)
FIG. 3.

lo-.

Assemble
Trim
generating bank f-- intemal piping
(Drums)
(Drums)

Alternative Component Networks for 'Drums'

3(b) shows another possible network for Drums, but it takes

longer to complete as pipefitters must work about 20 m [60
ft] in the air to assemble the generating bank and fit external
piping to the steam drum. Yet following this procedure may
be necessary if the length of the steam drum is greater than
the bay width of the boiler cavity (Le., the drum exceeds the
dimensions assumed for it by the designers of the structure),
so the drum must be lifted at an angle.
Product Network
A product network shows the sequence of construction of
all components in the product. It interlinks their component
networks. An "interlink" sequences activities of different
component networks, specifically:
1. A "network interlink" is a sequential link between two
component networks. It defaults to having all activities
of the preceding network (the network from which the
link emanates) precede with a FS link all activities of
the succeeding network (the network to which the link
points) but this can be customized by the user.
2. An "activity interlink" is a sequential link between two
activities of different component networks.
Activities, designated in CasePlan by a name-verb and a
name-noun (described later), e.g., "Erect Drums" can be consolidated if they are of the same type, Le., if they use the same
verb or verb phrase. For example, project specifications usually require conducting hydro tests for 'Drums', 'WW's, 'Superheater', and 'Economizer' when they are installed. Each
component must be tested, but testing each individually and
separately is meaningless as it does not provide any feedback
on the operation of the system as a whole. Therefore, all activities "Hydro test" are consolidated into a single one, to be
performed after all these components have been installed. Similarly, activities can be consolidated based on the type of their
associated components.
Activity
An activity has attributes to express the following:
1. Identification, namely: ID, identification number; nameverb: (e.g., "Erect") or verb phrase ("Trim external piping for") that describes the work to be performed;
name-noun: usually the component on which work is to
be done (e.g., 'Drums'); and name, which can be spec-

ified by the user directly or derived by CasePlan by combining the activity's attributes name-verb and name-noun
(e.g., "Trim external piping for Drums"). Derived
names are more amenable to reuse than user-specified
ones.
2. Sequencing, namely: predecessors, which lists the activities that immediately precede the activity, and successors, those that immediately succeed it.
3. Timing, namely scheduled times: the early and late start
and finish time, and four floats [Harris (1978) gives definitions and formulas]; and actual times: AS: actual start;
AF: actual finish; RD: remaining duration = duration (current date - AS); AD: actual duration = AF - AS.
Scheduled times are calculated using CPM based on the
sequencing and timing constraint attributes (described
next); they cannot be changed directly by the user. As a
minimum, the project start time, all activity durations,
and sequencing attributes need to be known for CasePlan
to calculate a schedule. The actual times become available only after the activity has started and possibly finished.
4. Timing constraints, namely: SNE (start no earlier than),
FNE (finish no earlier than), SNL (start no later than),
FNL (finish no later than), and others (Dzeng 1995).
They specify constraints that a contractor may impose on
an activity for a reason unrelated to network logic per
se, e.g., to reflect material delivery schedules or contract
milestones.
5. Resource allocation, namely to refer to materials (e.g.,
piping), laborers or crews (e.g., four pipefitters and a
foreman), or equipment (e.g., 30-ton PCSA class 12-105
crane) that are required to perform the activity. CasePlan
groups those resources into a 'Construction Method',
which is a class with attributes: crew, a list of individual
trades or crews and their number specified in the form:
[(crew-name crew-size) ...]; equipment, a list of equipment and its quantity specified in the form: [(equipmentname number-of-pieces-of-equipment) ...]; and productivity, the amount of work the specified crew and
equipment accomplish per day. Upon reuse of an activity,
the associated construction method is also reused, specifically, to estimate the activity duration based on the
productivity and possibly based on attributes of the associated component.
6. For-component and Use-condition (described later).

JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997/343

J. Constr. Eng. Manage. 1997.123:338-347.

Activity Link and Interlink between Component Networks

Each link or interlink specializes the Link class, which has
the following attributes:

Preceding-activity.
Succeeding-activity.
Type: SS, SF, FS, or FE
Lead-time, the amount of time between the link's head
and tail.
5. Strictness, quantified as a number between a and 1, with
a default value of 0.5. A "strictness threshold" (set at
0.8 but changeable by the user) determines whether a
link is strict (strictness 2: threshold) or not (strictness <
threshold). CasePlan uses strictness to gauge whether it
should reuse a link automatically (strict) or only after
user confirmation (not strict). A link that is strict reflects
a sequencing principle that applies to all projects, e.g.,
the sequential link between activities "Install columns
on floor 1" and' 'Install beams on floor 1" is quite strict
because one supports the other and is thus built first.
Conversely, a link that is not strict reflects that it was
introduced for a project-specific reason, e.g., a sequential
link between "Install east wall" and "Install north
wall" is not necessarily reusable because this ordering
may be the result of the availability of a specific crane.
6. Use-condition (described later).

Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

1.
2.
3.
4.

Relationships between Project and Schedule

CasePlan's user must annotate cases for archival by representing project knowledge and relating projects to their schedules. This is done using specific attributes and a language for
describing relationships.

For-Component and Use-Condition Attributes

Each part of a product network must have been included
for one of two reasons, which are specified using the following
attributes:
1. For-component, specified on classes 'Component-network', 'Activity', or 'Link', describes the relationship
between e.g., an activity ("Install drums") or network
and the component ('Drums') for which it represents a
construction alternative.
2. Use-condition, specified on classes 'Component-network', 'Activity', 'Link', 'Product-network', or 'Construction Method', expresses in a value specification language (described later) the conditions under which an
activity, link, network, or method can be reused. It relates
the class to constraints specified in the product of the
'Case', 'Site', or 'Project-Specifications'; e.g., the network for a 'Front WW' cannot start until the network
for 'Drums' is completed because the WW is connected
to the drums. Erecting the WW earlier would make the
connection work more difficult and the WW may get
damaged in the process. The Use-condition specifies that
the interlink between the two networks will be imposed
only if 'Front WW' has an is-connected-to relation with
'Drums'.

Value Specification Language

A "value specification" (VS), stored as value of an attribute, expresses how that value is to be computed and on what
project-specific data (instances of other attributes, activities,
links, components, methods, or products) it depends. It is written in CasePlan's VS language that includes functions, global
variables, and numerical values. Functions are those specific

to CasePlan, attribute names to retrieve values from attributes,

and all Common Lisp functions (Allegro 1994) except for constructive ones (Le., functions that build data structures, such
as defun or defclass). Scheduling knowledge is also expressed
using CasePlan's VS language. CasePlan recalculates each VS
each time the associated instance from an old case is reused
in a new project or when the new project is updated. Dzeng
(1995) provides examples.
A VS should be computable. It should avoid circular references and CasePlan must be able to evaluate it and yield a
result of the intended type (string, logic value, number, etc.)
for the associated attribute. A VS should also ease CasePlan's
reuse of activities and links that need adaption from one
project to another.

AUTOMATED SCHEDULING
Fig. 4 illustrates the steps CasePlan takes to automate the
planning and scheduling tasks: a double-edge rectangle represents a step for which CasePlan reuses cases; a single-edge
rectangle one for which it does not; a rounded rectangle represents information available before or after a step, and in it,
underlined text represents information generated or changed
by the preceding step.

Step 1: Determine Component Network

Given a new project, CasePlan first determines a network
for each product component by reusing the corresponding network (Le., the network used by a component of the same type)
of the best available case. CasePlan determines the best match
based on the component's similarity function. This yields a
"component similarity value" (CSV) with value between a
and 1, representing the similarity between a component in the
case being considered for reuse and the corresponding component in the new product. A CSV is an average with userassigned weights of the component's "attribute similarity values" (ASVs). An ASV is also a number between a and 1 but
it represents the similarity of a component's attribute in a case
to a corresponding attribute in the new product; e.g., "To what
extent is the Drums' total-mass in a Case similar to the Drums'
total-mass in the new product?" CasePlan uses several
schemes to measure the similarity of attribute values and to
normalize them according to the type of value that is being
compared (an example follows, but for more detail see Dzeng
1995).
A corresponding network will be reused if its usecondition-if specified-is satisfied in the new product setting. If so, the derived attributes of its activities and links (e.g.,
activity-duration) will be reused, but their values are not calculated at this time (they may depend on the network as a
whole). The associated construction method is also reused,
though it can be changed later. After completing this step,
information about component networks and construction methods and activity duration specifications are available (Fig. 4).
Suppose that CasePlan is to select a component network for
'Drums' of a new project (Project-I) based on two previous
cases (Case-A and Case-B). In Table 1, the Attribute column
lists the 'Drums' attributes with their user-assigned weights
(shown in parentheses) for finding the best case (higher
weights reflect greater importance). Columns "Project-IIAV,"
"Case-NAV," and "Case-B/AV" show the attribute values
for 'Drums' of Project-I, Case-A, and Case-B, respectively.
CasePlan computes the ASVs as indicated in columns
"Case-NASV" and "Case-B/ASV" using the user-specified
quantitative difference range for the attribute max-unit-mass:
[1.0 (0 4,000)], [0.8 (4,000 9,000)], [0.6 (9,000 20,000)],
[0.2 (20,000 30,000)], [0 (30,000 max.)]

344/ JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997

J. Constr. Eng. Manage. 1997.123:338-347.

Project
Cases

New case
Cases

t
Calculate
CPM
schedule

Determine
component
networks

Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

t
Project
Cases
Component networks
Interlinks
Construction methods
Activity durations

Project
Cases
Component networks
(Construction methods)
(Activity duration specs,)
\.

II'

Calculate derived
attribute values of
activities and links

Determine
product
network

Project
Cases
Component networks
Interlinks
(Construction methods)
(Activity duration specs,)

FIG. 4.

TABLE 1

total-mass (0.2)
max-unit-mass (0.2)
number-of-generating-tubes (0.4)
is-drum-1ength-greater-than-baywidth? (1.0)

Project-1
AV'
(2)

Case-A
AV
(3)

125,000 kg 90,000 kg 0.6

34,000 kg 30,000 kg 0.8
15,000
12,000
1.0
false

false

1.0

AV
(5)

Determine
construction
methods

CasePlan Subtasks

Case-B

ASV'
(4)

OR

Example Comparison or ProJect-1 and Two Cases

Attribute
(1)

Project
Cases
Component networks
Interlinks
Construction methods
(Activity duration specs,)

CSY (Case-B's 'Drums')

= 0.6

ASV
(6)

148,000 kg 0.6
45,000 kg 0.6
20,000
0.8
true

0.0

'Attribute values.
Attribute similarity values.

This specifies that the ASY is 1,0, 0.8, 0.6, 0.2, and 0,
respectively, if the difference of max-unit-mass of the project
and case being compared is between 0 (inclusive) and 4,000
kg, between 4,000 and 9,000 kg, between 9,000 and 20,000
kg, between 20,000 and 30,000 kg, and greater than 30,000
kg, respectively. Because the difference between the project
and Case-A is (34,000 kg - 30,000 kg) = 4,000 kg, Case-A's
ASY is 0.8; similarly, Case-B's ASY calculation yields 0.6.
Other ASYs are determined using different schemes but these
are not discussed here due to space limitations. The CSY of
'Drums' in each case is then calculated in the following two
equations:

0.2 + 0.6 X 0.2 + 0,8 X 0.4 + 0,0

0,2 + 0.2 + 0.4 + 1.0

1.0 = 0.31
(2)

Because Case-A has the higher CSY (0.93), its 'Drums'

network will be used for Project-l's 'Drums', provided that
no use-condition has been violated in the new project setting.
If all use-conditions are satisfied, CasePlan next examines each
individual activity and link of the network, and removes those
for which the use-condition has been violated. After each removal the user may have to relink disjoint activities. Networks
for other components are determined similarly.

Step 2: Determine Product Network

Second, CasePlan determines a product network by interlinking component networks and consolidating activities.
CasePlan chooses the best case for doing this based on the
product similarity value (PSY), which is an average with userassigned weights of the product's CSYs.
Suppose that the weights for determining a product network
are 1.0, 0.8, 0.8, 0,6, and 0,6 for components 'Boiler',
'Drums', 'Economizer', 'Front WW', and 'Rear WW', respectively, and Case-A's CSYs are 1.0, 0.93, 0.8, 1.0, and 1.0
for those same components, Then
PSV (Case-A)

CSY (Case-A's 'Drums')

= 0.6 x

1.0 X 1.0

0.2 + 0.8 x 0,2 + 1.0 x 0.4 + 1.0

0,2 + 0,2 + 0.4 + 1.0

1.0

= 0,93
(I)

= 0,943

0.8 X 0.93 + 0.8 X 0.8 + 0.6 X 1.0

1.0 + 0.8 + 0.8 + 0.6 + 0.6

0,6 X 1.0

(3)

JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997/345

J. Constr. Eng. Manage. 1997.123:338-347.

As a result of this step, activities of different component

networks are now interlinked.
Step 3: Determine Construction Method

Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

CasePlan can now reselect methods by reusing cases based

on component similarity functions with weights different from
those used for determining component networks. Alternatively,
it can do nothing and calculate activity durations based on
their known methods. Not all activities must have methods
specified; they need to only if their derived attributes refer to
the methods in one way or another, e.g., when activity duration
is a function of the method's productivity.
Step 4: Compute Schedule

CasePlan then calculates activity durations by evaluating the

VS of each activity's duration attribute (e.g., a heuristic factor
multiplied by the size of the activity's associated component,
or a formula based on the productivity of the selected method).
VSs for other derived attributes such as timing constraints are
also evaluated at this time, e.g., an activity that started no
earlier than May 20, 1991, being the completion date of the
steel structure of a boiler house, will start no earlier than December I, 1994, if it is reused in a new project where the
completion of the structure is on December 1, 1994.
Finally, CasePlan computes the remainder of the schedule
data using CPM. The schedule is then stored with the product
instance as a new case.
DECISION-SUPPORT SCHEDULING

CasePlan can function as a stand-alone or as a decisionsupport tool. Because CasePlan relates a product (and its components) to its schedule (and its component networks), a user
can browse and manually copy parts of existing schedules on
a component-by-component basis after choosing cases that are
similar to the project at hand. The user may modify component
networks or construction methods from selected cases or create
new ones from scratch. This allows for much more intelligent
assembly than is possible using "fragnets."
IMPLEMENTATION

CasePlan has been implemented in object-oriented lisp, Allegro CL/PC 2.0 (Allegro 1994), that runs in the Microsoft
Windows@ 3.1 environment. The model includes several
graphical display windows, examples of which are shown by
Dzeng (1995). The system includes a generic product model
for boilers of fossil-fueled power plants with a capacity ranging from 30 to 90 MW. The product model is the result of a
survey sent out to industry practitioners knowledgeable about
boiler erection and subsequently discussed with them face-toface or over the telephone. CasePlan's library currently includes seven realistic cases, including Grayling and Genesee,
with design details and preproject schedules obtained from two
boiler manufacturers.
CASEPLAN CAPABILITIES AND LIMITATIONS, AND
FUTURE RESEARCH

CasePlan makes selected knowledge from a project's paper

trail available to users by putting it on line. When successful
projects representing the best company practices are stored as
cases, CasePlan can help train schedulers new to those projects
by letting them browse through historic data, so they may
study why certain choices were made and experiment with
scheduling alternatives. Bad examples could also be stored but
should be annotated to warn the user when CasePlan retrieves
them.

CasePlan's use of a single standard to define a product and

its schedule eases the association between product components
and schedule subnetworks. It also allows CasePlan to develop
a new schedule by reusing subnetworks from different cases
instead of the entire network from a single case. Integrating
CasePlan with PlantSTEP and expanding the model describing
construction methods deserve further research attention.
Clearly, annotating cases and describing the relationships
between projects and schedules requires a considerable amount
of work. VSs are not unique and documenting the rationale
that was used to derive a certain schedule may be hard to come
by. Also, schedules pieced together from various sources and
"refreshed" with a new project's data may include contradictions that must be resolved manually before they can be
evaluated by CasePlan. Despite the effort it requires from the
user, CasePlan's VS language makes it possible to describe
schedule constraints, which could not be done using other
scheduling tools.
The requirement that scheduling knowledge be expressed in
a computable form may prevent CasePlan from using some
knowledge that human planners take into account. Nonetheless, it is worthwhile encoding what can be articulated as such,
since quite a few pieces of knowledge are computable, and if
there are many some may get overlooked by human planners.
CasePlan assumes that construction methods and activity
sequences in existing cases are satisfactory, that they can be
reused on an activity-by-activity basis, and that such reuse is
adequate for estimating activity durations. At this time
CasePlan has no way of accounting for the desirability of sharing or reusing resources in concurrent or consecutive activities.
Performing resource leveling for the schedule as a whole is
left to the user. Also, the question arose whether CasePlan
should use bid, preproject, or as-built schedules, but we found
pros and cons for each. Again, this issue is to be investigated
further.
The aim of CasePlan was to produce schedules that are field
executable. While CasePlan's architecture makes this possible,
we have been only moderately successful at capturing scheduling knowledge considered by the boiler erector. Boiler erectors derive their competitive advantage in part from knowing
their labor force well, and taking skill levels, motivation, and
learning-curve effects into account. The erector we interviewed declined to give out his expertise on this. This avenue
for research may be the most interesting one to pursue, as it
would increase our knowledge of how construction schedules
are broken down and used on site.
CasePlan follows an early-commitment approach, though it
has most of the mechanisms needed to generate multiple alternatives. We have not investigated how to keep track of multiple alternatives in a computationally efficient way, that is,
without having to spell out each individual alternative's details. An interesting research issue here is, 'Which cases
should be stored as individuals and which ones as parametrized variations of a prototype case?"
CasePlan uses the best it can find in the cases available to
it, so it will yield different results when supplied with different
cases. Users can add new cases and modify existing ones as
they see fit and expand the case base. This approach is fundamentally different from the one taken by existing AI-planners in which the planning knowledge is defined by the system
designer instead of the individual user of the system.
CONCLUSIONS

This paper presented CasePlan, a tool that uses case-based

reasoning to automate construction scheduling. It illustrates a
methodology to relate construction schedules to individual
projects and contractors' practice using product modeling, and
to reuse that knowledge in scheduling new projects. Cases are

346/ JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997

J. Constr. Eng. Manage. 1997.123:338-347.

annotated with scheduling knowledge to reflect key factors,

such as facility design considerations, site conditions, contractual requirements, contractor's practice, etc., that determined
why a schedule ended up being one way and not another.
CasePlan is a graphical decision-support tool that allows
schedulers to apply their own judgment in browsing and
copying reusable parts from schedules of previous cases whose
products are similar to the new one. This not only assists contractors in generating bid schedules when they are short of
time, but it also provides them with a means of documenting
and reapplying their company's best practices.

Downloaded from ascelibrary.org by National Chiao Tung University on 05/01/14. Copyright ASCE. For personal use only; all rights reserved.

ACKNOWLEDGMENTS
We would like to acknowledge the kindest support from and participation in this research by industry practitioners at Asea Brown Boveri
Combustion Engineering Services, Inc., Black & Veatch, Inc., The
Christman Co., Genesee Power Station, Grayling Generating Station,
Northern Boiler Mechanical Contractors, Inc., Townsend & Bottum, Inc.,
and Zurn Industries, Inc.
This research was funded by grant MSS-9215935 from the National
Science Foundation (NSF), whose support we gratefully acknowledge.
Any opinions, findings, conclusions, or recommendations expressed in
this paper are those of the authors and do not necessarily reflect the views
of NSF.

APPENDIX.

REFERENCES

Allegro CUPC 2.0, object oriented development system for Windows.

(1994). Franz, Inc., Berkeley, Calif.
"Biomass-fired plant practices total resource management." (1993).
Power Engrg., (April), 33-36.
Cherneff, J., Logcher, R., and Sriram, D. (1991). "Integrating CAD with
construction-schedule generation." J. Compo in Civ. Engrg., ASCE,
5(1),64-84.
"Construction industry: OSHA safety and health standards." (1985). 29
CFR 1926/1910, OSHA 2207, Occupational Safety and Health Administration, Washington, D.C.
Darwiche, A., Levitt, R. E., and Hayes-Roth, B. (1989). "OARPLAN:
generating project plans by reasoning about objects, actions and resources." AI EDAM, 2(3), 169-181.
Desmond, C. (1995). "New Genesee Twp. power plant to be fueled by
clean waste wood." Mich. Contractor & Builder: The Weekly Jobs
Book, Contractor Publishing Co., Detroit, Mich., (Jan. 21), 6-8.
Dzeng, R. J. (1995). "CasePlan: a case-based planner and scheduler for

construction using product modeling," PhD Dissertation, Civ. and Envir. Engrg. Dept., Univ. of Michigan, Ann Arbor, Mich.
Dzeng, R. J., and Tommelein, I. D. (1993). "Using product models to
plan construction." Proc., 5th Int. Con! Compo in Civ. and Build.
Engrg., ASCE, New York, N.Y., 1778-1785.
Dzeng, R. J., and Tommelein, 1. D. (1994). "Case storage of planning
knowledge for power plant construction." Proc., 1st Congr. Compo in
Civ. Engrg., ASCE, New York, N.Y., 293-300.
Echeverry, D., Ibbs, w., and Simon, K. (1991). "Sequencing knowledge
for construction scheduling." J. Constr. Engrg. and Mgmt., ASCE,
117(1), 118-130.
Gray, C. (1986). "Intelligent construction and cost analysis." Constr.
mgmt. and economics, London, U.K., 4, 135-150.
"Guidelines for specifying integrated computer-aided engineering applications for electric power plants." (1987). EPRI NP-5159M, Electric
Power Res. Inst., Palo Alto, Calif.
Harris, R. B. (1978). Precedence and arrow networking techniques in
construction. John Wiley & Sons, New York, N.Y.
ISO Standard 10303-1. (1994a). "Product data representation and
exchange. Part I: overview and fundamental principles." Int. Standards
Org.
ISO Standard 10303 AP. (1994b). "Project proposal for plant spatial configuration AP." Int. Standards Org.
Kartam, N. A., and Levitt, R. E. (1990). "Intelligent planning of construction projects." J. Compo in Civ. Engrg., ASCE, 4(2), 155-176.
Kolodner, J. (1993). Case-based reasoning. Morgan Kaufmann, San Mateo, Calif.
Miresco, E. (1992). "Expert system for scheduling building projects."
Proc., CIB 92 Vol. 1, Nat. Res. Council Canada, 258-259.
Morad, A. A., and Beliveau, Y. J. (1991). "Knowledge-based planning
system." J. Constr. Engrg. and Mgmt., ASCE, 117(1), 1-12.
Navinchandra, D., Sriram, D., and Logcher, R. (1988). "GHOST: project
network generator." J. Compo in Civ. Engrg., ASCE, 2(3), 239-254.
Primavera project planner for Windows 1.0. (1993). Primavera Systems
Inc., Ba1a Cynwyd, Pa.
Special issue on data and product modeling. (1996). J. Camp. in Civ.
Engrg., ASCE, 10(3).
Thabet, W. Y., and Beliveau, Y. J. (1993). "A model to quantify work
space availability for space constrained scheduling within a CAD environment." Proc., 5th Int. Can! Camp. Civ. and Build. Engrg., ASCE,
New York, N.Y., 110-116.
Tommelein, 1. D., Levitt, R. E., Hayes-Roth, B., and Confrey, T. (1991).
"SightPlan experiments: alternate strategies for site layout design." J.
Camp. in Civ. Engrg., ASCE, 5(1), 42-63.
Zozaya-Gorostiza, C., Hendrickson, C., and Rehak, D. R. (1989). Knowledge-based process planning for construction and manufacturing. Academic Press, Boston, Mass.

JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT / SEPTEMBER 1997/347

J. Constr. Eng. Manage. 1997.123:338-347.