NEURAL-SLIDING CONTROL WITH CMAC

IN VARIABLE STRUCTURE SYSTEMS

Celso P. Bottura
UNICAMP - Machines, Components and Intelligent Systems Department
P.O. Box 6101 - 13081-970 - Campinas, SP, Brazil
Marcelo C. M. Teixeira
UNESP - Electric Engineering Department
Av. Brasil, 56 - Ilha Solteira, SP 15385-000, Brazil
Maurício J. Bordon
UNICAMP - Machines, Components and Intelligent Systems Department
P.O. Box 6101 - 13081-970 - Campinas, SP, Brazil

Abstract. This work proposes a new method of dynamical control systems

implementation using the Cerebellar Model Articulation Controller (CMAC)
neural network as part of a variable structure system. CMAC is a neural
network easily implemented in hardware, quite efficient in the reproduction of
multivariable functions and extremely fast in its training process. These
features yield CMAC applications in robust multivariable on line control in the
presence of uncertainties in plant model. Variable Structure Systems (VSS)
theory yields the concatenation of at least two, possibly unstable, dynamical
structures, in order to realize a low cost controller that guarantees robustness
with respect to nonidealities in the control system. This is done by a deliberate
introduction of a special behavior named sliding mode where the state of the
plant is confined to a subspace determined by the eigenvalues of the desired
resultant dynamics until the origin is reached. Further, if the system model
equations are written in a particular regular form and complete access to
states is guaranteed for any time, the CMAC network in a variable structure
system collects information enough about plant dynamics for the closed loop
eigenvalues be allocated on line with no previous knowledge of plant model
other than its order and structure. Furthermore, if the supposed unknown plant
parameters change due to environment interaction, the resultant dynamics is
invariant.

Keywords: CMAC, Variable structure systems, Sliding mode control.

1. INTRODUCTION

The actual growth of works on variable structure control systems with

sliding mode (Ackermann & Utkin, 1998) is due to their quasi invariance
property to parameter disturbances (Chen & Chang, 1996), to the extension to
the multivariable control problems and to the attractive simplicity of the relay
type control usually applied to the plant. The variable structure control
systems can reach desired dynamic behavior and show stability certainty as a
feature but, usually, the knowledge of the linear system model matrices and
is a need for sliding mode controller design, (Furuta, 1990) and (Iordanou
& Surgenor 1997). This article proposes a method for variable structure
control of partially known and/or uncertain dynamical systems where a CMAC
neural network approximates a quantity depending on the system state and on
the system parameter structure. The experimental results show that the
approximation obtained is accurated enough to yield the system state to reach
the discontinuity surface where it evolves in sliding mode to the origin.

2. CEREBELLAR MODEL ARTICULATION CONTROLLER - CMAC

Since its conception in the mid-seventies (Albus, 1975), the Cerebellar

Model Articulation Controller (CMAC) has been used in on line nonlinear
control systems, including robotics and chemical processes (Brown & Harris,
1994). Also, applications in reinforcement learning control, state estimation
(Wang, Brown & Harris, 1994), digital filtering and color calibration hardware
realization (Ker, Kuo & Wen, 1997) are reported. CMAC approximates
multivariable functions over a compact domain of interest only (Ker, Kuo &
Wen, 1997), and in order to improve accuracy, higher order basis functions
than the binary one can be used. Memory requirements is a critical issue in
using CMAC due to its exponential growth with accuracy requirements
(Bordon, 1995). Albus (1975) suggested hashing code to randomly attribute
one weight to many association cells.Teixeira & Bordon(1995) used discrete
in value weights to reduce the memory needs in function approximation.
The CMAC output, (Hirashima & Iiguni, 1997) and (Tolle & Ersü, 1992),
is the sum of weights, where is a positive integer, chosen from a set
with weights, by quantization functions applied to the set of
inputs, where is a compact set of Each quantization function
determines a weight associated with one interval of the partition in that
contains the input vector to the CMAC. In all, quantizations are made and
there exists an overlapping of of an interval in each coordinate axis. In
Fig.1, the set is partitioned by quantization functions. The first
one associates one weight to each interval in the set . The
other weight is associated with one of the intervals of the
second quantization function. The CMAC output is defined in Eq.(1):




where are the weights chosen by the quantization functions for some
For two input vectors and sharing weights determined by the
quantization functions, the corresponding outputs and are similar, and,
in this context, is a measure of the generalization inherent to the CMAC.
We define the learning rule for CMAC in Eq.(2):



is the desired value for the output associated to some , are

the weights whose summation results in and is a positive constant.
Usually

3. VARIABLE STRUCTURE SYSTEMS

This section collects some aspects of Variable Structure Systems (VSS)

Theory for multivariable linear systems that lead to a nice fusion between
sliding mode control and CMAC based on line training control. Let be the
null space of a linear transformation , defined by




is the intersection of surfaces in and, in multivariable VSS

control, it is called sliding subspace (Utkin, 1992).

 
   ! " #%$& ('*),+-'/.

 Let the time invariant multivariable linear system written in a particular
regular form be:

0

We suppose that the state is accessible but the matrix is not completely
known. Let's consider the following partitioning of matrix and :

21

where is an identity matrix and the eigenvalues of are due only to

and ( and could represent a priori known relations about state
variables).
Let the discontinuous control law be:

23

where is the th component of the vector and and

are continuous functions with respect to and .
The procedure for stabilizing this system may be divided in two phases
(Utkin, 1978):
· to yield the state, initially equal to , to a subspace in the state
space called sliding subspace:
4

where is a matrix of chosen accordingly to desired conditions for the

dynamic behavior during the next phase;
· to keep the system state in a sliding domain of the discontinuity surface
where a fast switching process, named sliding mode, is established due to
discontinuous control law (6).
The system state remains in sliding mode in the subspace if there
exists a function satisfying the following conditions, (Utkin, 1978)
and (Utkin, 1992):
 8


Once the state reaches the sliding subspace and remains on it, with
a switching frequency as high as desired, the resultant dynamics becomes
independent of the plant parameters and depends only on . So, if the sliding
mode is established and the control system parameters satisfy the existence
conditions for sliding, the closed loop dynamics is independent of the
parameters of the plant.
Let a standard quadratic Lyapunov function be

25

Consider the matrix is partitioned as in Eq.(10)

6
7

where stands for the identity matrix. When the sliding regime is established,
we have and

6



in addition, the sliding mode for the system (3) in a neighborhood of the
subspace exists if:

6


th
where is the row of matrix .

Discrete Sliding Mode. Unlike the continuous time VSS, where changes
in the structure due to switching may occur at any instant, in a computer
controlled system, a quasi-sliding regime appears due to switching changes
only at sampling instants. An alternative approach to approximate the sliding
regime that occur in a digital control procedure to the ideal case is increasing
sampling frequency; see in Furuta, (1990). In this way, the structure changing
delay will be limited to the sampling period.
On an ideal sliding situation, the switching frequency tends to infinity and
apparently the state remains on the subspace Equivalent control method
(Utkin, 1978) may then be applied to study the properties of the continuous
resultant dynamics. But in a real sliding situation, the state moves in a finite
neighborhood of the subspace and, so,

6
9


Utkin (1978) showed that, for a certain class of systems, the real sliding
occurs in a neighborhood close to the trajectory described by the ideal sliding.

4. ALGORITHM DESCRIPTION AND RESULTS

The proposed controller was applied in the stabilization of a partially

unknown plant, with access to all states, and given by

6
0

Figure 2 shows the proposed controller performance. In this work,

each partition divides the coordinate axes in 30 intervals, and the
CMAC is trained at each . Figure 3 and 4 show the control law and
the value of during simulation. The simulated time was and values
of taken in the interval showed to be satisfactory. The CMAC
used in this application has three inputs related to the components of the
vector which represents the state of the system. The control law employed
is defined by

6
1

where is a diagonal matrix and

6
3

CMAC output and training are given by:

6
4

6
98

 :    ;&< >=<?  @= ;&AB$  C = %$<DE$ .  F@G< B$<H   .

where is a matrix chosen accordingly to desired dynamic properties for the

closed loop system; is an approximation to obtained by digital
differentiation, for example; is a diagonal matrix that controls the sliding
robustness. Supposing the state is far away the origin, the quantity
guarantees the validity of the existence conditions (12) if is not close
enough to Initially, the state is far away the origin, the CMAC
output is not adjusted to yield the state in the sliding surface direction and the
norm of may increase. At the end of some learning iterations,
approximates the neighborhood that contains the state becomes a sliding
domain, for the existence conditions are now verified, and the sliding starts
at the first crossing of the discontinuity surface. According to existence
conditions, to yield the sliding mode, the control law (6) must be such that

: I' $  JA ? K$<A ? H%$< ? ELMAK  K$ .  B$<H @G<

6
5

and, in this application,

7

As the CMAC output, , is intended to imitate , if

the existence conditions for sliding are satisfied and the robustness is
guaranteed.
   0  NPO A%  $Q >=  O B G A@  O C  ELR .  B$<H >G< .

5. CONCLUSIONS

A procedure for neural-sliding controller realization with CMAC and VSS

is proposed. Due to robustness properties of sliding mode behavior, this
approach yields the stabilization of the controlled process to a quasi-invariant
regime, and this is done with a chattering free control action in the
stabilization point due to nonlinear state feedback employed. For a measure of
the tolerance of the controller to variations in plant parameters it is necessary
a deeper theoretical study. Further studies include discrete time systems
analysis which yields lower switching/sampling frequencies.

REFERENCES

Ackermann, J. & Utkin, V., 1998, Sliding mode control design based on
Ackermann’s formula, IEEE Transactions on Automatic Control, vol.43,
n.2, pp. 234-237.
Albus, J. S., 1975, A new approach to manipulator control: the cerebellar
model articulation controller (CMAC), Transactions of the ASME Journal
of Dynamical Systems Measurements and Control, vol.97, n. 9, pp. 220-
 227.
Bordon, M. J., 1995, Sistemas de controle automático baseado no cerebellar
model articulation controller, Ms.C. Thesis, Engineering School of Ilha
Solteira - UNESP, Ilha Solteira, SP, Brazil.
Brown, M. & Harris, C. J., 1994, Neurofuzzy adaptive modelling and control,
Prentice Hall, New York.
Chen, F. C., & Chang, C. H., 1996, Practical stability issues in CMAC neural
network control systems, IEEE Transactions on Control Systems
Technology, vol. 4, n. 1, pp. 86-91.
Furuta, K., 1990, Sliding mode control of a discrete system, System & control
letters, vol. 14, pp. 145-152.
Hirashima, Y. & Iiguni, Y., 1997, An identification system design using a
CMAC with a learning algorithm based on the Kalman filter. Proceedings
of the 11 th IFAC Symposium on System Identification, July 8-11,
Kitakyushu, Fukuoka, Japan, vol. 3, 1389-1394.
Iordanou, H. N., & Surgenor, B. W., 1997, Experimental evaluation of
robustness of discrete sliding mode versus linear quadratic control, IEEE
Transactions on control systems technology, vol. 5, n. 2, pp. 254-260.
Ker, J., Kuo, Y., Wen, R. & Liu, B., 1997, Hardware implementation of
CMAC neural network with reduced storage requirement, IEEE
Transactions on Neural Networks, vol. 8, n. 6, pp. 1545-1556.
Teixeira, M. C. M., & Bordon, M. J., 1995, Improving memory requirements
using cerebellar model articulation controller with discrete weights,
Proceedings of the VI International Fuzzy Systems Association World
Congress, July 22-28, São Paulo, vol. II, pp. 599-602.
Tolle, H. & Ersü, E., 1992, Neurocontrol: learning control systems inspired by
neuronal architectures and human problem solving strategies, eds. M.
Thoma and A. Wyner, Springer-Verlag, New York.
Utkin, V. I., 1978, Sliding modes and their application in variable structure
systems, Mir, Moscow.
Utkin, V. I., 1992, Sliding modes in control optimization, Springer-Verlag,
Berlim.
Wang, H., Brown, M. & Harris, C. J., 1994, Neural network modelling of
unknown nonlinear systems subject to immeasurable disturbances, IEE
Proceedings on Control Theory Applications, vol. 141, n. 4, pp. 216-222.