MULTIPLE MODELING AND CONTROL OF NONLINEAR SYSTEMS

WITH SELF-ORGANIZING MAPS

By
JEONGHO CHO

A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL

OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
2004

Copyright 2004
by
JEONGHO CHO

This document is dedicated to my late father My inspiration and the one who gave me
every opportunity to realize my dreams.

ACKNOWLEDGMENTS
Being at CNEL has been not only a wonderful academic experience but also a
unique opportunity to meet many colleagues who helped me immensely along the way.
It is a privilege to me to acknowledge the unconditional support of my supervisor,
Dr. Jose C. Principe, who has been a mentor during my years as a graduate student. His
advice, wisdom, and many invaluable lessons in life have made this dissertation possible.
I sincerely appreciate the help offered by the members of my academic committee,
Dr. John G. Harris, Dr. Michael C. Nechyba, and Dr. Loc Vu-Quoc. Their cooperation
and suggestions have considerable improved the quality of my dissertation. I would also
like to express my gratitude to Dr. Mark A. Motter and Dr. Deniz Erdogmus for their
help and providing many valuable comments during the past years of my stay in the
CNEL.
My deepest recognition goes to my beloved parents, especially to my late father
who helped me in any imaginable way to achieve my objectives and fulfill my dreams.
They have been an inexhaustible source of love and inspiration all my life.
My most special thanks go to my wife, Joonhee, for her patience, understanding
and encouragement, without which it would have been impossible to complete this
dissertation. Finally, I thank my specially beloved son, Minsuh, hoping that the effort of
these years may offer him a more plentiful life in the years to come.

iv

TABLE OF CONTENTS
page
ACKNOWLEDGMENTS ................................................................................................. iv
LIST OF TABLES............................................................................................................ vii
LIST OF FIGURES ......................................................................................................... viii
ABSTRACT....................................................................................................................... xi
CHAPTER
1

INTRODUCTION ........................................................................................................1
1.1 Motivations .............................................................................................................1
1.2 Review of Literature ...............................................................................................2
1.3 Objectives and Authors Contribution....................................................................5
1.4 Outline ....................................................................................................................6

SYSTEM IDENTIFICATION VIA MULTIPLE MODELS .......................................8

2.1 Local Dynamic Modeling .......................................................................................8
2.2 SOM-Based Local Modeling................................................................................10
2.2.1 Reconstruction of State-Space....................................................................11
2.2.1.1 Delay reconstruction ........................................................................11
2.2.1.2 Estimating an embedding dimension ...............................................12
2.2.2 The Self-Organizing Map...........................................................................14
2.2.2.1 Competitive process .........................................................................15
2.2.2.2 Cooperative process .........................................................................16
2.2.2.3 Adaptive process ..............................................................................17
2.2.3 Modeling Methodology Based on the SOM...............................................17
2.3 Input-Output Representation of Systems..............................................................19
2.3.1 Classical Approach.....................................................................................20
2.3.2 Series-Parallel and Parallel Models............................................................21
2.3.3 SOM-based Multiple ARX Models............................................................23
2.3.3.1 Selection of operating regions with a SOM .....................................23
2.3.3.2 Model development procedure.........................................................27

MULTIPLE MODEL BASED CONTROL ...............................................................29

3.1 Discrete-Time Control System .............................................................................31

3.2 Inverse Control via Backpropagation Through Model.........................................32
3.3 Multiple Inverse Control.......................................................................................34
3.4 Multiple PID Control............................................................................................37
4

MULTIPLE QUASI-SLIDING MODE CONTROL .................................................42

4.1 Introduction to Variable Structure Systems..........................................................42
4.1.1 Sliding Hyperplane Design.........................................................................44
4.1.2 Sliding Mode Control Law Design.............................................................45
4.2 Sliding Mode Control in Sampled-Data Systems.................................................48
4.2.1 Quasi-Sliding Mode....................................................................................49
4.2.2 Quasi-Sliding Mode Control Using Multiple Models ................................52
4.3 Analysis of Multiple Quasi-Sliding Mode Control with an Imperfect Sensor .....54

CASE STUDIES.........................................................................................................57
5.1 Controlled Chaotic Systems .................................................................................57
5.1.1 The Lorenz System.....................................................................................58
5.1.2 The Duffing Oscillator ...............................................................................66
5.2 Nonlinear Discrete-Time Systems ........................................................................72
5.2.1 A First-order Plant......................................................................................72
5.2.2 A Laboratory-scale Liquid-level Plant .......................................................78
5.3 Flight Vehicles......................................................................................................86
5.3.1 Missile Dynamics .......................................................................................86
5.3.2 LoFLYTE UAV..........................................................................................92

CONCLUSIONS AND FUTURE WORK...............................................................101

6.1 Summary.............................................................................................................101
6.2 Future Work........................................................................................................103
6.3 Concluding Remarks ..........................................................................................104

APPENDIX
QSMC FOR MIMO SYSTEM ........................................................................................106
LIST OF REFERENCES.................................................................................................109
BIOGRAPHICAL SKETCH ...........................................................................................118

vi

LIST OF TABLES
Table

page

5-1

Lipschitz index of the controlled Lorenz system for determining an embedding

dimension. ................................................................................................................59

5-2

Comparison of modeling performance for the controlled Lorenz system. ..............62

5-3

Comparison of modeling performance for the controlled Duffing oscillator. .........67

5-4

Comparison of tracking performance for 3 different control task : Settling time

and NRMS-SSE.........................................................................................................70

5-5

Lipschitz index of a laboratory-scale liquid-level plant for determining an

embedding dimension. .............................................................................................79

5-6

Comparison of modeling performance for the liquid-level plant.............................80

5-7

Comparison of control performance for the liquid-level plant in noise-free

environment..............................................................................................................83

5-8

Comparison of control performance for the liquid-level plant in the presence of

sensor noise: standard deviation of noise is 4.5e-2. .................................................83

5-9

Comparison of modeling performance for the missile system.................................87

5-10 Comparison of controller performance for the missile system in the presence of
noise having the standard deviation of 4.6e-2..........................................................90
5-11 Comparison of modeling performance for the lateral motion ( p and r ) of the
LoFLYTE UAV. ......................................................................................................95

vii

LIST OF FIGURES
Figure

page

2-1

Local modeling scheme on the basis of a SOM. ......................................................10

2-2

Two data points that are close in 1 but distant in 2 .............................................13

2-3

Kohonens Self-organizing Map. .............................................................................15

2-4

Nonlinear dynamic model configuration..................................................................20

2-5

A series-parallel model (left) and A parallel model (right)......................................22

2-6

Configuration of local linear modeling based on a SOM.........................................27

3-1

Classical discrete-time control system. ....................................................................31

3-2

Modeling and control scheme using the TDNN: (a) TDNN modeling of a plant
(b) An inverse controller via Backpropagation through (Plant) Model. ..................34

3-3

Proposed SOM-based inverse control scheme. ........................................................36

3-4

PID controlled system. .............................................................................................38

3-5

Overall schematic diagram of the nonlinear PID closed loop control mechanism
using multiple controllers.........................................................................................39

3-6

Block diagram of PID controller for a SISO plant model........................................40

4-1

Phase plane plot of a continuous-time second-order variable structure system.......46

4-2

Discrete-time system response with sliding mode control.......................................49

5-1

The uncontrolled Lorenz system: phase-space trajectory and time-series. ..............59

5-2

Generalization error v.s. Number of PEs (left) Learning curve (right)....................60

5-3

Identification of the controlled Lorenz system by multiple models.........................61

5-4

Tracking a fixed point reference signal by MIC: (a) y d = 0 (b) y d = 8 . ...............63

viii

5-5

Comparison of control performance varying the number of inverse controllers

based on multiple models. ........................................................................................64

5-6

Comparison of tracking performance by multiple model based controllers (MIC,

MPIDC, MQSMC) and global inverse controllers (IC-ARX, TDNNC). ................65

5-7

The uncontrolled Duffing oscillator: phase-space trajectory and time-series..........66

5-8

Lipschitz index (left) for the determination of optimal number of inputs and
outputs and Generalization error v.s. Number of PEs (right)...................................67

5-9

Identification of the controlled Duffing oscillator by TDNN (left) and multiplemodels (right). ..........................................................................................................68

5-10 Performance comparison on trajectory tracking by TDNNC, PID-ARX, and

MPIDC when the poles of the closed-loop response are place at (a) 0.9 (b)
0.5 0.5i (c) 0.25 0.25i (d) 0.05 0.05i. ...............................................................69
5-11 Control performance by TDNNC (left) and MPIDC (right). ...................................71
5-12 Parameter selection to design multiple models. .......................................................72
5-13 Modeling performance using 64 multiple models for a nonlinear first
order plant. ...............................................................................................................73
5-14 Responses for parameter selection to design QSMC by varying
(a) rT and (b) qT . .................................................................................................74
5-15 Comparison of tracking performance using a global controller and multiple model
based controllers in the absence of sensor noise. The figure (right) is an
enlargement of the figure (left) between 34 and 52 iterations. ................................75
5-16 Performance of square-wave tracking in the absence of noise by the MQSMC......76
5-17 Comparison of performance against noise among TDNNC, MIC, MPIDC and
MQSMC. .................................................................................................................77
5-18 Sinusoidal and arbitrary signal tracking by the MQSMC:(a) in the absence of
sensor noise (b) in the presence of sensor noise, SNR = 20dB . ..............................78
5-19 Modeling performance using multiple models for a liquid-level plant....................79
5-20 Typical input-output characteristic of the second-order liquid-level plant. .............81
5-21 Square-wave tracking performance of the liquid-level plant varying the sliding
surface and the noise level by the MQSMC.............................................................81

ix

5-22 Control of the liquid-level plant by the MQSMC varying the number of
controllers: (a) M = 1 (b) M = 16 (c) M = 36 (d) M = 144.......................................82
5-23 Control of the liquid level system with measurement noise by the MQSMC
with (a) M = 1 (b) M = 16 (c) M = 144 and (d) the TDNNC. ..................................84
5-24 Tracking an oscillatory reference signal of the liquid-level plant by the TDNNC
(left) and the MQSMC (right) in the presence of sensor noise. ...............................85
5-25 Performance assessment on a trajectory tracking under noisy environment. ..........85
5-26 Modeling performance using multiple models for the missile dynamics. ...............87
5-27 Tracking various set-point reference signal by (a) TDNNC (b) MIC (c) MPIDC
and (d) MQSMC in the absence of measurement noise...........................................88
5-28 Tracking various set-point reference signal by (a) TDNNC (b) MIC (c) MPIDC
and (d) MQSMC under the presence of measurement noise whose standard
deviation is 4.6e-2. ...................................................................................................89
5-29 Trajectory tracking by TDNNC (left) and MQSMC (right) in the presence of
noise whose standard deviation is 2.3e-1. ................................................................90
5-30 Set-point tracking behavior by the TDNNC (left) and the MQSMC (right) under
parameter variations. ................................................................................................91
5-31 General description of aircraft (left) and LoFLYTE testbed UAV (right).............92
5-32 Control inputs ( a , r ) used to generate data samples for training the networks. ...94
5-33 Modeling of a roll-rate using multiple models.........................................................95
5-34 Comparison for controlling roll-rate and yaw-rate to track the set point in the
absence of noise by (a) TDNNC (b) MIC (c) MQSMC...........................................98
5-35 Comparison for controlling roll-rate and yaw-rate to track the set point in the
presence of noise by (a) TDNNC (b) MIC (c) MQSMC. ........................................99
5-36 Performance of controlling roll-rate and yaw-rate to track an arbitrary trajectory
with measurement noise (SNR = 20) by (a) TDNNC (b) MIC (c) MQSMC.........100

Abstract of Dissertation Presented to the Graduate School

of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy
MULTIPLE MODELING AND CONTROL OF NONLINEAR SYSTEMS
WITH SELF-ORGANIZING MAPS
By
Jeongho Cho
December 2004
Chair: Jose C. Principe
Major Department: Electrical and Computer Engineering
This dissertation is concerned with the development and analysis of a nonlinear
approach to modeling and control of nonlinear complex systems. In particular, the
problem of designing a mathematical model of a nonlinear plant using only observed data
is considered.
For the identification of the plants, the concept of multiple models with switching
is employed in order to simplify both the modeling and the controller design since a
single controller may sometimes have difficulty meeting the design specifications in case
the dynamics vary considerably over the operating region. For this reason, a SelfOrganizing Map (SOM) is utilized to divide the operating region into local regions as a
modeling infrastructure to construct local models. The SOM selects the local operating
region relying on the embedded output, and the local model is built by the embedded
output as well as the embedded control input data samples which are spaced in the local
area.

xi

Based on the identified multiple models, the problem of designing controllers is

discussed. Each local linear model is associated with a linear controller, which is easy to
design. Switching of the controllers is done synchronously with the active local linear
model that tracks different operating conditions. The effectiveness of the proposed
approach is shown through experiments for modeling complex nonlinear plants such as
chaotic systems, nonlinear discrete time systems and flight vehicles. Its comparison with
neural networks-based alternatives, Time Delay Neural Network (TDNN), shows clear
advantages of local modeling and control in terms of performance.

xii

CHAPTER 1
INTRODUCTION
1.1 Motivations
The identification of nonlinear dynamical systems has received considerable
attention since it is an indispensable step towards analysis, simulation, prediction,
monitoring, diagnosis, and controller design for nonlinear systems [21,65,67]. In
particular, the problem of designing a mathematical model of a nonlinear plant using only
observed data has attracted much interest, both from an academic and an industrial point
of view. During the past few years, neural networks as a global model have been
suggested for nonlinear dynamical black-box modeling and successfully applied to the
prediction and modeling of nonlinear processes [11,46,65,73].
Global models, however, have shown some difficulties in cases when the
dynamical system characteristics vary considerably over the operating regime, effectively
bringing the issue of time varying parameters (or nonlinearity) into the design. On the
other hand, local modeling derives a model based on neighboring samples in the
operating space to characterize some operating point or similar feature [23,85]. If a
function f to be modeled is complicated, there is no guarantee that any given global
representation will approximate f equally across all space. Moreover, nonlinear models
are too complex to be used for controller design [70]. Thus, nonlinear control methods
cannot serve all needs of real industrial control problems.
In this case, the dependence on representation can be reduced using local
approximation where the domain of f is divided into local regions and a separate model is
1

2
used for each region [3,13,24]. In a number of local modeling applications, a SelfOrganizing Map (SOM) has been utilized to divide the operating regions into local
regions [30,60,74,97]. The SOM is particularly appropriate for switching, because it
converts complex, nonlinear statistical relationships of high-dimensional data into simple
geometric relationships that preserve the topology in the feature space [44]. Thus the role
of the SOM is to discover patterns in the high dimensional state space and divide that
space into a set of regions represented by the weights of each Processing Element (PE).
Linear models and associated techniques for linear control design are typically used
to control the plant under certain specific operating conditions. This type of control is
only valid in a small region around the operating point. For that reason, the concept of
multiple models with switching, according to a change in dynamics, has been an area of
interest in control theory in order to simplify both modeling and controller design
[62,66]. The motivation for this research, therefore, is to explore control strategy using
SOM-based multiple models for nonautonomous and nonlinear systems.
1.2 Review of Literature
There are many examples in the literature in which the local modeling paradigm
has been successfully applied for the modeling of nonlinear autonomous and
nonautonomous systems. Farmer and Sidorowich [23] have shown that local linear
models, despite their simplicity, provide an effective and accurate approximation of
chaotic dynamical systems. Jacobs et al. [38] have proposed the mixtures of expert
models that are composed of several different expert networks and a gating network that
localizes the experts. They showed that a simple model can be built by dividing a vowel
discrimination task into appropriate subtasks. Bottou and Vapnik [3] have proposed using
local learning algorithms instead of training a complex system with all data samples, and

3
demonstrated that a set of subsystems trained with a subset of data can improve the
performance for an optical character recognition problem. The neural-gas architecture
proposed by Martinetz et al. [53] is similar to the SOM in that the competitive network
divides the input space in a set of smaller regions and then local linear models are created
by a LMS-like rule. They showed that the neural-gas network outperforms MLPs and
RBF networks for time series prediction. The same group [77] used a SOM for the
control of a robotic arm. Murray-Smith et al. [61] similarly have extended RBF networks
where each local model is a linear function of the input and exhibited great success in
control problems. Principe and Wang [74] have successfully modeled a chaotic system
with a SOM-based local linear modeling method. Vesanto [94] and Moshou and Ramon
[60] proposed a scheme that essentially followed local linear modeling based on SOM
topology for nonautonomous system. Under some conditions, it has been shown that
multiple models can uniformly approximate any system on a closed subset of the state
space provided a sufficient number of local models are given.
Generally, the control using multiple models is categorized by two approaches: a
global model-based control using local models and a multiple model-based control with
switching. Global controller design with the aid of multiple linear models has been
extensively reported in the literature. Gain scheduling has been perhaps the most
common systematic approach to control nonlinear systems in practice due to its simple
design and tuning [47,68,79]. The multiple model adaptive control approach differs from
gain scheduling mainly in the use of an estimator-based scheduling algorithm used to
weight the local controllers. Murray-Smith and Hunt [61] utilized an extended RBF
network where each local model is a linear function of the input and reported great

4
success for control problems. The overall controller designed is based on the local models
and a validity function to guarantee smooth interpolation. Similarly, Foss et al. [24] and
Gawthrop and Ronco [29] employed model predictive controllers and self-tuning
predictive controllers, respectively, using multiple models. Palizban et al. [70] attempted
to control nonlinear systems with the linear quadratic optimal control technique using
multiple linear models and provided the stability condition for the closed loop system.
Ishigame et al. [37] proposed the sliding mode control scheme based on fuzzy modeling
composed of a weighted average of linear systems to stabilize an electric power system.
In contrast, Narendra et al. [66] proposed the multiple model approach in the
context of adaptive control with switching where local model performance indices have
been used to select the local controller. Subsequently, Narendra and Balakrishnan [64]
proposed different switching and tuning schemes for adaptive control that combine fixed
and adaptive models yielding a fast and accurate response. Principe et al. [75] proposed a
SOM-based local linear modeling strategy and predictive multiple model switching
controller to control a wind tunnel and showed improved performance with decreased
control effort over both the existing controller and an expert human-in-the-loop control.
Later, Narendra and Xiang [63] proved that the adaptive control using multiple models is
globally stable and that the tracking error converges to zero in the deterministic case.
Diao and Passino [20] applied multiple model based adaptive schemes to the fault
tolerant engine control problem. A linear robust adaptive controller and multiple
nonlinear neural network based adaptive controllers were exploited by Chen and
Narendra [11]. Thampi et al. [89,90] have also shown the applicability of the multiple
model approach based on the SOM for flight control.

5
1.3 Objectives and Authors Contribution
This dissertation is concerned with modeling and control of nonlinear
nonautonomous dynamical systems. The objective of this dissertation is to investigate if
it is possible to obtain a better result in extending the formulation of the control problem
from using just one global model to using several internal models. Thus a multiple
modeling approach is presented and techniques to design controllers based on these
model structures are developed.
The main contributions made by the author with respect to the modeling and
control of nonlinear systems include:
Firstly, an extended version of the SOM-based local modeling scheme for
nonautonomous and nonlinear plants is developed for more general representation of the
underlying dynamical systems and better approximation solely based on input-output
measurements of the plant. Local linear models are derived through competition using the
SOM and they are derived from the data samples corresponding to each of the SOMs
PEs.
Secondly, we investigate several options regarding how to capture the dynamics in
the input-output joint space. It is shown that as the number of dependent variables is
increased SOM modeling may become increasingly difficult to model accurately due to
its memory based structure. Thus, the SOM is trained to position the local models in the
embedded output space. At any time instant, the model representing the plant dynamics is
chosen by the SOM depending on the history of the plant and then incorporated with the
previous control inputs.
Thirdly, the model structure is controller oriented since the dynamics are simpler
locally than globally such that it is easier to develop local models as well as controllers.

6
For instance, if the system phenomena or behavior changes smoothly with the operating
point, then a linear model (or controller) will always be sufficiently accurate locally
provided that the operating region is sufficiently small, even though the system may
contain complex nonlinearities when viewed globally. These local controllers can then be
switched as the system changes operating conditions. Hence, multiple control with
switching, such as an inverse controller and a PID controller using identified multiple
models, are examined.
Finally, in order to obtain a controller which preserves the good sensitivity to
external disturbances, a sliding mode controller (SMC) is employed using multiple
models. By doing so, one of the difficulties in designing a SMC (that requires the
complete knowledge of the plant to be controlled) can be removed. In addition, we
examine the effect by the modeling error due to the quantization of state space as well as
by measurement noise to the proposed multiple model based sliding mode control
performance. It is shown that the switching scheme does not create an issue to be
considered in order to guarantee BIBO stability of the overall system.
1.4 Outline
This dissertation is divided into six chapters. Chapter 2 gives a review of the local
dynamic modeling required to study further for nonautonomous systems. The SOM
employed as a modeling infrastructure to construct the local models is described briefly.
At the end, the proposed SOM-based multiple ARX modeling scheme is introduced for
nonlinear nonautonomous dynamic systems representation. Chapter 3 shows how to
design an inverse control and a PID control framework based on the designed multiple
linear models. A description of a global nonlinear TDNN trained by backpropagtion
through a model is also presented for comparisons regarding performance. Chapter 4

7
gives a brief description of Variable Structure Systems (VSS). Quasi-sliding mode
control strategy is proposed based on multiple models. In addition, analysis of multiple
quasi-sliding mode control structure with an imperfect sensor is discussed. In Chapter 5,
simulations are conducted assuming that both the plant is unknown and the only state
available for measurements is the plant output for two controlled chaotic systems, two
nonlinear discrete-time systems, one missile, and one Unmanned Aerial Vehicle (UAV).
Finally, Chapter 6 presents conclusions based on the preceding analysis and simulation
results and suggests further study.

CHAPTER 2
SYSTEM IDENTIFICATION VIA MULTIPLE MODELS
The idea of multiple modeling is to approximate a nonlinear system with a set of
relatively simple local models valid in certain operating regimes [39]. Because of the
complexity, uncertainty and nonlinearity of a large class of systems, we often cannot
derive appropriate models from first principles, and are not capable of deriving accurate
and complete equations for input-state-output representations of the systems. Hence we
need to resort to input-output data in order to derive the unknown nonlinear system model
[10,35]. The technique of multiple model networks is appealing for modeling complex
nonlinear systems due to its intrinsic simplicity [62,66].
2.1 Local Dynamic Modeling
We begin with a brief overview of a dynamical systems approach to input-output
modeling. When no physical knowledge of the system is available, we have to determine
a model from a finite number of measurements of the systems inputs and outputs. An
autonomous dynamical systems approach to black-box modeling based on Takens
Embedding theorem was first suggested by Casdagli [7]. The delay embedding offers the
possibility of accessing linear or nonlinear coupling between variables and is a
fundamental tool in nonlinear system identification. The use of delay variables in the
structure of these dynamical models is similar to that originally studied by Leontaritis and
Billings [49], and is common in linear time-series analysis and system identification [96].
When we are trying to understand an irregular sequence of measurements, an
immediate question is what kind of process generates such a series. Under the
8

9
deterministic assumption, irregularity can be autonomously generated by the nonlinearity
of the intrinsic dynamics. Let the possible state x of a system be represented by points in
a finite dimensional phase space, P . This can be realized by a map of P onto itself:
x k +1
M =

x k P + 2

xk

f M
x

k P +1

(2.1)

The predictive mapping is the centerpiece of modeling since once determined, f can be
obtained from the predictive mapping f i : P as
x k +1 = f i ( x ,k )

(2.2)

where x ,k = [ xk , x k 1 , L, xk P +1 ]T . In addition, Singer et al. [85] derived the locally linear

prediction based on this relationship as
rT
f i ( x ,k ) a i x ,k + bi

(2.3)

r
The vector and scalar quantities of a and b are estimated from the selected pairs

( x j +1 , x , j ) in the least square sense, where j is the index of the data samples in the
operating regime, i.e., one model. To obtain a stable solution, more than P pairs must be
selected. In general, the above local model fitting is composed of two steps: a set of
nearby state searches over the signal history and model parameters which, when pieced
together, provide a global modeling of the dynamics in state space. The underlying
dynamics f is then approximated as

fi

(2.4)

i =1,L, N

where N is the number of operating regimes. Based on this approximation of an

autonomous system, local linear models have performed very well in comparative studies

10
on time series prediction problems and in most cases have generated more accurate
predictions than global methods [84,97]. Moreover, the nonlinear dynamical system can
be identified by local framework even in the presence of noise if enough data are
available to cover all of the state space since local regions are local averages of the data.
To make the local network less sensitive to noise and outliers, more than one neighbor
can be utilized in local modeling.
2.2 SOM-Based Local Modeling
The SOM is employed as a modeling infrastructure to construct the local models. It
provides a codebook representation of the plant dynamics and organizes the different
dynamic regimes in topological neighborhoods. Thus we can create a set of models that
are local to the data in the Voronoi tessellation created by the SOM. This local model
structure with the SOM is depicted in Figure 2-1.
uk

yk

Plant
Reconstruction of State-Space

y k +1

Model 1
Model 2
M

Switching
Device
(SOM)

Model N

Figure 2-1. Local modeling scheme on the basis of a SOM.

11
2.2.1 Reconstruction of State-Space
In many cases of practical interest it is not possible to measure the state variables of
r
a system directly. Instead, the measuring procedure yields some value y k = ( x k ) , when
r
the system is in states x k . Here, () is a measurement function which in general

depends on the state variables in a nonlinear way. The time evolution of the state of the
system results in a scalar time series x1 , x 2 , x 3 , L . In order to reconstruct the underlying
dynamics in phase space, delay embedding techniques are commonly used.
2.2.1.1 Delay reconstruction
Delay-coordinate embedding [41,91], a technique developed by the dynamics
community, is one way to help the input-output modeling; it allows one to reconstruct the
internal dynamics of a complicated nonlinear system from a single time series. That is,
one can often use delay-coordinate embedding to infer useful information about internal
(and immeasurable) states using only output information. The reconstruction produced by
delay-coordinate embedding is not, of course, completely equivalent to the internal
dynamics in all situations, or embedding would amount to a general solution to control
the theorys observer problem: how to identify all of the internal state variables of a
system and assume their values from the signals can be observed [87]. However, a singlesensor reconstruction, if done properly, can still be extremely useful because its results
are guaranteed to be topologically (i.e., qualitatively) identical to the internal dynamics.
This means that conclusions drawn about the reconstructed dynamics are also true of the
internal dynamics of the system inside the black box. In order to reconstruct the
underlying dynamics in phase space, we begin with scalar observable, y k of the state x k

12
of a deterministic dynamical system. Then typically we can reconstruct a copy of the
original system by considering blocks

y ,k = [ ( x k ), ( x k ),L, ( x k ( dy 1) )]T
r

= y k , y k , L , y k ( dy 1)

(2.5)

of dy successive observations of , for dy sufficiently large. Delay-coordinate

reconstruction is governed by two parameters, embedding dimension dy and embedding
delay . Note that using dy = 1 merely returns the original time series; one-dimensional
embedding is equivalent to not embedding at all. Proper choice of dy and is critical to
this type of phase-space reconstruction and must therefore be done wisely; only correct
values of these two parameters yield embeddings that are guaranteed by the Takens
theorem [88] and subsequent work by Packard et al. [69] and Casdagli et al. [7] to be
topologically equivalent to the original (unobserved) phase-space dynamics.
2.2.1.2 Estimating an embedding dimension
There has been much work on determining the embedding dimensions of the time
series generated by autonomous dynamical systems in the absence of dynamical noise
[41,91,42,6]. The methods developed for estimating the minimum embedding dimensions
are grounded on Takens embedding theorem [88] and most of them use the ideas of the
false nearest neighbors technique [42,6]. Later a number of works discussed theoretical
foundations of the delay embedding of the input-output time series [7,6]. This led to the
generalization of the existing method for the case of non-autonomous dynamical systems
[87,6,76].
He and Asada [35] proposed a strategy which is based directly on measurement
data and does not make any assumptions about the intended model architecture or

13
structure. It requires only that the process behavior can be described by a smooth
function, which is an assumption that must be made in black box nonlinear system
identification. An explanation of this strategys central idea follows. In general case, the
task is to determine the relevant inputs of the function
y k +1 = f (1 , 2 ,L , n )

from a set of potential inputs 1 , 2 ,L , o

(2.6)

(o > n) that is given. If the function in (2.6)

is assumed to depend on only n 1 inputs although it actually depends on n inputs, the

data set may contain two (or more) points that are very close (in the extreme case they
can be identical) in the space spanned by the n 1 inputs but differ significantly in the
nth input. This situation is shown in Figure 2-2 for the case n = 2 .

2
yi
yj

1
Figure 2-2. Two data points that are close in 1 but distant in 2 .
The two points i and j are close in the input space spanned by 1 alone but they are
distant in the 1 2 input space. Because these points are very close in the space
spanned by the n 1 inputs ( 1 ) it can be expected that the associated process outputs yi
and y j are also close (assuming that the function f () is smooth). If one (or several)
relevant inputs are missing then obviously yi and y j are expected to take totally

14
different values. In this case, it is possible to conclude that the n 1 inputs are not
sufficient. Thus, the nth input should be included and the investigation may begin again.
In [35] an index is defined based on so-called Lipschitz quotients, which is large if
one or several inputs are missing (the larger the quotients, the more inputs are missing)
and is small otherwise. Thus, using this Lipschitz index the correct embedding
dimensions can be detected at the point where the Lipschitz index ceases to decrease. The
Lipschitz quotients in the one-dimensional case are defined as

lij =

yi y j

i j

for i, j L , i j

(2.7)

where L is the number of samples in the data set. For the multidimensional case, the
Lipschitz quotients can be calculated by the straightforward extension of (2.7):
lijn =

yi y j
(1,i 1, j ) + L + ( n ,i n , j )
2

for i, j L , i j

(2.8)

where n is the number of input. The Lipschitz index, then, can be defined as the
maximum occurring Lipschitz quotient

l n = max (l ijn )
i , j ,( i j )

(2.9)

As long as n is too small and thus not all relevant inputs are included, the Lipschitz
index will be large because situations as shown in Figure 2-2 will occur. As soon as all
relevant inputs are included, (2.9) stays relatively constant.
2.2.2 The Self-Organizing Map
The principal goal of the SOM is to transform an incoming signal pattern of
arbitrary dimension into a one or two-dimensional discrete map, and to perform this
transformation adaptively in a topologically ordered fashion [44]. Figure 2-3 shows

15
Kohonens model of a two-dimensional SOM. Each PE in the lattice is fully connected to
all the source PEs in the input layer. This network represents a feedforward structure with
a single computational layer consisting of PEs arranged in rows and columns.

Winning PE

x1,k
x 2,k
x m ,k
Input Layer

Competition Layer

Figure 2-3. Kohonens Self-organizing Map.

The algorithm responsible for the formation of the SOM proceeds first by
initializing the synaptic weights in the network. Once the network has been properly
initialized, there are three essential processes involved in the formation of the SOM:
competition, cooperation and adaptation. Descriptions of these processes follow.
2.2.2.1 Competitive process
Let m denote the dimension of the input space. Let an input vector selected
randomly from the input space be denoted by
r
T
x k = [x1,k , x 2,k , L, x m,k ]

(2.10)

The synaptic weight vector of each PE in the network has the same dimension as in the
input space. Let the synaptic weight vector of PE i be denoted by
r
wi = [ w1,i , w2,i , L , wm ,i ], i N

where N is the total number of PEs in the network.

(2.11)

16
r
r
To find the best match of the input vector x k with the synaptic weight vectors wi ,
r
r
compare the Euclidean distance between x k and wi and select the smallest one as
r
r
i o = argmin x k wi , i = 1, L , N

(2.12)

which sums up the essence of the competition process among the PEs. According to
(2.12), i o is the subject of attention because we want the identity of PE i. The particular
PE i that satisfies this condition is called the best-matching unit or winning PE for the
r
input vector x k .

2.2.2.2 Cooperative process

The winning PE locates the center of a topological neighborhood of cooperating
PEs. A topological neighborhood can be defined by many methods. In particular, a PE
that is firing tends to excite the PEs in its immediate neighborhood more than those
farther away from it, which is intuitively satisfying. This means that after classification of
the input sample, the adaptation will be done not only for the winning PE but also for the
neighbors of the PE which gives the best response. Let i o ,i denote the topological
neighborhood function centered on the winning PE i o , then a typical choice of i o ,i is

i o ,i , k

r ro
i i
= exp
2
2 k

(2.13)

where ri ri o represents the Euclidean distance in the output space between the ith PE
and the winning PE and k is the effective width of the topological neighborhood. To
satisfy the requirement that the size of the topological neighborhood shrinks with time, let
the width of the topological neighborhood function decrease with time

17
k
, k = 1,2, L
1

k = o exp

(2.14)

where o is the value of at the initiation of the SOM algorithm, and 1 is a time
constant. Thus, as time (i.e., the number of iterations) increases, the width decreases at an
exponential rate, and the topological neighborhood shrinks in a corresponding manner.
2.2.2.3 Adaptive process
The network can be trained with a simple Hebbian-like rule to train the weights of
the winning PE and its neighbors. The neighboring PEs can be trained in proportion to
their activity (Gaussian), or all neighbors within a certain distance can be trained equally.
The learning rule can be described as follows:
r
r
r
o
r
wi ,k + k i ,k ( x k wi ,k ), i = i
wi ,k +1 = r
otherwise
wi ,k ,

(2.15)

Notice that both the learning rate, k , and neighborhood size, k , are time dependent
and are typically annealed (from large to small) to provide the best performance with the
smallest training time.
2.2.3 Modeling Methodology Based on the SOM
In this architecture of local linear modeling, the SOM is trained to position the local

models in the embedded output space, y ,k = y k , y k 1 , L, y k ( dy 1)

. The SOM

preserves topological relationships in the input space in such a way that neighboring
inputs are mapped to neighboring PEs in the map space. Then, when each PE is extended
with a local model it can actually learn the mapping y k +1 = f ( y ,k ) in a supervised way.

r
Each PE has an associated local model { ai , bi } in (2.3) that represents the approximation
of the local dynamics.

18

r
The local model weights { ai , bi } are computed directly from the desired signal
samples y i , j and the input samples by a least square fit within a Voronoi region centered
at the current winning PE chosen from y . The size of the data samples in the region
must be at least equal to the dy -dimensional basis vector. The design procedure for this
local model is as follows:
1.

Apply training data to the SOM and find the winning PE corresponding to the input
y such that we have winner-input pairs.

2.

Use the least square fit to find the local linear model coefficients for the winning
PE, i o , where desired output vector y i , j M as

o
r
y i o , j = [ aiTo bi o ] y ,i , j for j M
(2.16)
1
r
r
where = [ a iTo bi o ] is the sought linear model coefficients, M is the size of inputs
involved in the winning PE i o .
Specifically, the least-squares problem
r
Y = X

(2.17)

r
is solved for , where X ( dy M ) is defined as a matrix that contains each input vector

associated with the winning PE, and Y M is defined as a vector that contains the
target outputs. It is well-known that although the least-squares solution obtained from
(2.17) is reasonably good when the noise level is low, the estimates tend to be biased for
higher levels of noise. Addition of a single sample to a cluster can radically change the
distances. Besides, the models will perform very well for that particular training set with
very low error because it has memorized the training examples but they may not perform
well with new data sets. Thus we make use of data samples from the winner as well as
the neighbors to create the local models in order to make them more robust as well as to

19
improve the generalization for the network. Also we take the data samples from the
neighbors in case less data than the dimension of the input are assigned in some Voronoi
region.
In testing, once the winning PE is determined we select the appropriate local model
from the list of associated models. Apply the local model to obtain the estimated output
r
y k +1 = a iTo y , k + bi o

(2.18)

2.3 Input-Output Representation of Systems

The temporal state evolution of an autonomous system is functionally dependent
only on the system state, but a nonautonomous system, such as considered in this work,
allows for an explicit dependence on an independent variable, the control input, in
addition to the system state. For an autonomous system, it is reasonable to assume that
the future behavior of the system can be predicted over some finite interval from a finite
number of observations of past outputs. In contrast, predictions of the behavior of a
nonautonomous system require consideration of not only the internal deterministic
dynamics (past outputs), but also of the external driving term (future input) [30,75,96].
System identification is a technique that permits building mathematical models of
dynamic systems based on input-output data (measurements). Its main purpose is to
identify a model of an unknown process in order to predict and gain insight into the
behavior of the process [39]. Real-life systems almost always show nonlinear dynamical
behavior. This behavior complicates the task of finding models that accurately describe
these systems. While in a large number of applications a linear model shows already
satisfactory results, there are numerous situations where linear models are not accurate
enough; especially when we deal with very complex systems or require very high

20
performance. Physical knowledge of the system can be a great aid in finding a nonlinear
model. However, this knowledge is not always available. In these cases we have to
determine a model from a finite number of measurements of the systems inputs and
outputs. This approach to nonlinear system modeling is often referred to as nonlinear
black-box identification. Usually, a nonlinear mapping is fitted from a number of delayed
inputs and outputs to the current output [94]. This results in a nonlinear input-output
model of the system.
2.3.1 Classical Approach
Some common classical approaches for nonlinear nonautonomous system modeling
are based on polynomials, e.g., Kolmogorov-Gabor polynomial models [71], Volterra
Series models [110], Hammerstein models [16,22,33], and so on, for the realization of the
nonlinear mapping.

uk

z 1

z 1

z 1

z 1

z 1

z 1

yk

Nonlinear static approximator f ()

y k +1
Figure 2-4. Nonlinear dynamic model configuration.
Normally, a discrete-time nonlinear dynamic system can be described by a NARX
(Nonlinear Auto-Regressive with eXogenous input) model that is an extension of the
linear ARX model, and represents the system by a nonlinear mapping of past inputs and
output terms to future outputs, that is,

y k +1 = f ( y k , L, y k dy +1 , u k ,L, u k du +1 )

(2.19)

21
Here y k Y p is the output vector and u k U q is the input vector. For
simplicity, we will set p = q = 1 . Let the (dy + du) - dimensional basis vector be

k = [ y ,k , u ,k ]T = [ y k ,L, y k dy +1 , u k ,L, u k du +1 ]T

(2.20)

where k is in the set = Y dy U du . Figure 2-4 shows the schematic diagram of NARX
model. Another nonlinear model is a NOE (Nonlinear Output Error) structure described
by
y k +1 = f ( y k , L, y k dy +1 , u k ,L, u k du +1 )

(2.21)

where y is the output of the identification model f . A NARMAX (Nonlinear AutoRegressive Moving Average with eXogenous input) is the lagged version of the NARX
model and is represented by
y k +1 = f ( y k , L, y k dy +1 , u k ,L, u k du +1 ) + ek

(2.22)

In the above, f can be replaced by neural networks, radial basis function networks or
fuzzy logic systems, which are other methods that have been developed for nonlinear
system identification [5,96]. Narendra and Parthasarathy [65] have compared NOE and
NARX, and as a result they have shown that NARX is better than NOE. In the neural
network community most identification schemes use the series-parallel model (NARX).
2.3.2 Series-Parallel and Parallel Models
A nonlinear dynamic model can be used in two configurations: a series-parallel
model and a parallel model. A series-parallel model predicts one or several steps into
the future on the basis of previous plant inputs and plant outputs and ensures that all the
signals are bounded if the plant is BIBO stable. Most published reports use the seriesparallel model because of its resulting stability. A requirement for using this model is that

22
the plant output is measured during the operation. In particular, in control engineering
applications the series-parallel model plays an important role, e.g., for the design of a
minimum variance or a predictive controller.

uk

yk

Plant , P( z )
z 1

z 1 z 1

Model, f ()

z 1

uk

yk

Plant , P( z )
z 1

z 1 z 1

Model, f ()

y k +1

z 1
z 1

y k +1

Figure 2-5. A series-parallel model (left) and A parallel model (right).

In contrast, a parallel model is required whenever the plant output cannot be
measured during operation. This is the case when a plant is to be simulated without
coupling to the real system, or when a sensor is to be replaced by a model. Also, for fault
detection and diagnosis the plant output may be compared with the simulated model
output in order to extract information from the residuals. Finally parallel model is very
useful when dealing with noisy systems since it avoids problems of bias caused by noise
on the real system output: If the identification model is to be used offline, the parallel
model is obviously more suitable. The parallel model, however, lacks theoretical
verification; hence, it is difficult to utilize its advantages.
The two configurations shown in Figure 2-5 can not only be distinguished for the
model operation phase but also during training. In this research, we follow the seriesparallel model.

23
2.3.3 SOM-based Multiple ARX Models
In the interest of modeling the local dynamics of a nonautonomous system in each
region, the local approximation method presented for autonomous systems can be
extended by letting x ,k = k in (2.3), so that y k +1 = f i ( k ) . Provided that necessary
smoothness conditions on f i : Y are satisfied, a Taylor series expansion can be
used around the operating point. The first-order approximation about the systems
equilibrium point produces N local predictive ARX models f 1 , L , f N of the plant
described by
f i ( k )

dy 1

du 1

j =0

j =0

ai, j y k j + bi, j u k j , i = 1,L, N

(2.23)

where ai , j and bi , j are the parameters of the ith model. Although higher order Taylor
approximations would improve accuracy, they are not very useful in practice because the
number of parameters in the model increases drastically with the expansion order.
Our proposed methodology is summarized as follows: first, the delayed version of
input-output joint space is decomposed into a set of operating regimes that are assumed to
cover the full operating space1. Next, for each operating regime we choose a simple linear
ARX model to capture the dynamics of the region. Consequently, a nonlinear
nonautonomous system is approximated by a concatenation of local linear models
f ( )

f i ( y , u )

(2.24)

i =1,L, N

2.3.3.1 Selection of operating regions with a SOM

Building local mappings in the full operating space is a time and memory
consuming process, which led to the natural idea of quantizing the operating regimes and
building local mappings in positions given by prototype vectors obtained from running

24
the plant. For quantization of the operating regimes, the k-nearest-neighbor method is
effective but it disregards neighborhood relations, which may affect performance [53]. In
contrast, the SOM has the characteristic of being a local framework liable to limit the
interference phenomenon and to preserve the topology of the data using neighborhood
links between PEs. Neighboring PEs in the network compete with each other by means of
mutual lateral interactions, and develop adaptively into specific detectors of different
signal patterns [44]. The training algorithm is simple, robust to missing values, and it is
easy to visualize the map. These properties make SOM a prominent tool in data mining
[94].
In most of the papers discussing local linear models for system identification, the
SOM has been used with a first order expansion around each PE in the output space. The
SOM transforms an incoming signal pattern of arbitrary dimension into a one or twodimensional discrete map, and performs this transformation adaptively in a topologically
ordered fashion [44]. The results obtained so far with this methodology have been quite
promising. However, problems that need to be solved remain: first, efficiently
partitioning the operating regimes in high dimensional spaces is still a problem due to the
curse of dimensionality [30]; second, it may be hard to find a small number of variables
to characterize the operating regimes due to the possibly large number of local models;
third, all the methods have to be extended for nonautonomous regimes.
The previous work by Principe et al. [75] provided the starting point for the
proposed modeling architecture. The most important difference is how to capture the
dynamics in the input-output joint space, which is fundamental for identifying the
unknown system. Several options are possible, and we have been investigating them:

25
Firstly, we tried to find the local models by quantizing the input-output joint space
by embedding not only the outputs but also the control inputs using one SOM. This
modification is essential because the purpose is to characterize the system dynamics that
exist in the input-output joint space. However, we encountered some difficulties such as
normalization of the joint space and large dimensionality of the space involved (many
degrees of freedom and large dynamic range of parameters) [13].
Secondly, in order to reduce the approximation error with local models based on a
SOM, we utilized a counter-propagation network by quantizing the input-output joint
space and the desired signal space together [15]. Since the output at each PE is just the
average output for all of the feature vectors that map to that point local models might be
created for better approximation using the quantization error in the input space and the
average output. This is achieved by coupling each PE with a linear mapping in such a
way that a functional relationship can be established between each Voronoi region in the
input space (of the SOM) and the desired signal. However, this method required a much
larger map to make the estimation error in the desired output space smaller. Additionally,
when noise is added in the input of the SOM, the quantization error in the input Voronoi
region may be magnified by the local models.
As the number of dependent variables is increased, the process becomes
increasingly difficult to model accurately. This led us to think that a model that uses only
a few of the observed variables will be more accurate than a model that uses all the
observed variables. In this scheme, therefore, we let the SOM look at only the current
output and its past values to decide the winner, and create the models with the control
inputs.

26
Here we will pursue the last option for the following reasons. The competitive
learning rule works best for normalized inputs. The SOM algorithm uses the Euclidean
metric to measure distances between vectors. For example, if one variable has values in
the range of [-100,,100] and another in the range of [-1,,1] the former almost
completely dominates the map organization because of its greater impact on the measured
distances. Either, the measure of distance is weighted by the inverse of the scales or the
data must be normalized such that each component of the input vectors have unit variance
and zero means [8]. However, normalization loses information (the mean or the scale can
be important) and it can become meaningless if the data dynamic range (or mean)
changes over time. Therefore we cannot normalize the data (nor create the weighted
Euclidean metric) in this way since it is not always guaranteed that the mean and the
dynamics range of the data are available.
In addition, as the number of dependent variables is increased, SOM modeling
becomes increasingly difficult because it is basically a memory-based approach that does
not scale up well with the input dimension. This led us to think that a model that uses
only a few of the observed variables will be more accurate than a model that uses all of
the observed variables. When the SOM modeling is done in the output space, we let the
SOM look at only the current output and its past values to decide the winner which
represents the operating regime, and create the models with the control inputs as shown
in Figure 2-6. In so doing, normalization of the input space is not necessary since the
clustering is performed solely by the history of the output.

27
uk

yk

Plant

y,k

Embedding

u,k

1
0

r
b1
r
b2

Embedding

SOM

r
a1

r
a2

y k +1

r
aN

r
bN

Figure 2-6. Configuration of local linear modeling based on a SOM.

2.3.3.2 Model development procedure
After the operating regions are divided by the SOM the underlying dynamics f is
then approximated as f U iN=1 f i , where N is the number of operating regions. N local
predictive ARX models f 1 , L , f N of the plant are described by
dy

du

j =0

j =0

f i ( k ) ai , j y k j + bi , j u k j , i = 1,L, N

(2.25)

where ai, j and bi, j are the parameters of the ith model. Then, when each PE of the SOM is
extended with a local model it can actually learn the mapping y k +1 = f i ( y ,k , u ,k ) in a
supervised way. The development of local models is done by directly fitting the
quantized embedded output samples obtained from the SOM and corresponding
embedded control input samples that cover the whole range of operation of the plant.

28

r r
Each PE has an associated local model { a i , bi } which are computed directly from
the desired signal samples ri , j and the input-output samples by a least square fit within a
Voronoi region centered at the current winning PE chosen from y . The design
procedure for this local model is as follows:
1.

Apply training data to the SOM and find the winning PE corresponding to the input
y such that we have winner-input pairs.

2.

Use the least square fit to find the local linear model coefficients for the winning
PE, i o , where desired output vector ri o , j M as
r r o
ri o , j = aiTo biTo y ,i , j
u ,i o , j

for j M

(2.26)

r r
where [ aiTo biTo ] is the sought linear model coefficients, M is the size of data

involved in the winning PE i o .

3.

In testing, once the winning PE is determined we select the appropriate local model
from the list of associated models. Apply the local model to obtain the estimated
output
r
r
y k +1 = aiTo y ,k + biTo u ,k
(2.27)
Our proposed modeling methodology is summarized as follows: first, the delayed

version of input-output joint space is decomposed into a set of operating regions that are
assumed to cover the full operating space. Next, for each operating region we choose a
simple linear ARX model to capture the dynamics of the region. Consequently, a
nonlinear nonautonomous system is approximated by a concatenation of local linear
models.

CHAPTER 3
MULTIPLE MODEL BASED CONTROL
Researchers have been interested in control of nonlinear systems for a very long
time. Progress in nonlinear control design, however, has been difficult because of the
intrinsic complexity of the problem [82]. In general, nonlinear control methods are
complex and can be applied only to a narrow class of systems. For example, methods
such as backstepping and feedback linearization can be applied to nonlinear systems with
some specific structure, but not to arbitrary nonlinear systems. Thus, nonlinear control
methods cannot serve all needs of real industrial control problems.
One way to approach the control of a nonlinear system in a wide range of
conditions is to linearize the model at a number of operating points, and then design one
linear feedback controller at each operating region [101]. These local controllers can then
be switched or scheduled as the system changes operating conditions. The use of
multiple models is not novel in control theory. Multiple Kalman filters were proposed in
the 1960s and 1970s by Magill [51] and Lainiotis [45] to improve the accuracy of the
state estimates in control problems. Fault detection and control in aircraft was proposed
by Maybeck and Pogoda [54], and in the subsequent years Maybeck and Stevens [55]
used the idea extensively in controlling aircraft systems. In all the above cases, no
switching is concerned, and only a linear combination of the control determined by the
different models is used to control the system.
The idea of switching between controllers has been most likely introduced for the
first time in the adaptive control literature by Martensson [52]. In the direct switching
29

30
schemes, the sequence in which the different controllers are to be tried is pre-determined.
The only determination that has to be made is when to switch from one controller to
another. It was soon realized that such architectures have very little practical utility. On
the other hand, the outputs of the multiple observers determine both when and to which
controller switching should occur in indirect switching schemes. Middleton et al. [56]
explicitly proposed the use of multiple models and switching to alleviate the problem of
stabilizing of the estimated model in indirect control, and further extended in [59] and
labeled the Hysteresis switching algorithm. The objective in all the above efforts is to
attain stability in adaptive control with minimum past information.
A controller is often highly dependent on a plant model especially when the
controller has been designed out of the model. Hence, for those cases, the modeling error
would be a relevant criterion for controller selection. If the number of controllers is
bounded, the delay between the selection of the controller and its activation can be
neglected. Thus the selection of the controller according to the modeling error is feasible.
Narendra and Balakrishnan [64] were the first to propose this idea of using multiple
adaptive models and switching in order to improve the performance of an adaptive
system, while assuring stability. Although it has already been shown that the performance
of a system can be significantly improved using the multiple model adaptive control with
switching, applicability to highly complex systems with this approach has not been
investigated in details. Thus, in this chapter, multiple controller design methodologies are
introduced by extending the multiple model approach for more complex nonlinear
systems.

31
3.1 Discrete-Time Control System
The field of classical control theory concerns itself with the task of servo or
regulator control of linear analog plant. Design methods for both continuous-time linear
controllers and discrete-time linear controllers obtained by discretizing the plant are well
understood. Figure 3-1 shows a schematic diagram of a classical discrete-time control
system.
dk
rk +1

Controller
G( z)

uk

Plant
P( z )

y k +1

z 1

Figure 3-1. Classical discrete-time control system.

The signal rk +1 is the reference signal. We would like the plant output y k +1 to track
it as closely as possible. To track the reference signal, the controller uses both rk +1 and
y k to compute the plant control signal u k . Feedback of y k is used to stabilize the plant,

and to ensure that the controller is both resilient in the face of external disturbances and
able to quickly reduce the output error to zero. Their only drawback is that they assume
precise knowledge of the plant dynamics. For this reason, a great deal of effort has been
expended to create accurate models of typical plants. As one improved way of modeling,
we proposed multiple models for better approximation of the plant in Chapter 2. This
scheme makes it easier to design the controller for the model which approximates internal
descriptions of the plant when given a finite number of external measurements without
any knowledge of the plant dynamics.

32
One reason to consider discrete-time systems is that it is well known that most
complex systems are controlled by computers which are discrete in nature and this
constitutes an obvious reason for dealing with multiple model based control. Another is
the fact that the presence of random noise can be dealt with more easily in the case of
discrete-time systems. Since most practical systems have to operate in the presence of
noise, the stability and performance of multiple model based control in such contexts has
to be well understood, if the theory is to find wide applications in practice.
3.2 Inverse Control via Backpropagation Through Model
In order to design a controller, we need to determine a plant model which should
capture the dynamics of the plant well enough that a controller designed to control the
plant model will also control the plant very well. Such a model might be derived from
physics by carefully analyzing the system and determining a set of partial differential
equations which explain its dynamics. Alternatively, the model might be a black-box
implementing some sort of universal transfer function. This function may be tuned by the
adjustment of its internal parameters to capture the dynamics of the system. For nonlinear
unknown systems, a NARX model of sufficient order is a universal dynamic system
approximator. Hence, we implement NARX neural network plant models for
performance comparisons with the proposed local linear models.
The conventional design methods for control systems involve constructing a
mathematical model of the systems dynamics and utilization of analytical techniques for
the derivation of a control law. Such mathematical models comprise sets of linear or
nonlinear differential/difference equations, which are usually derived with a degree of
simplification or approximation. The modeling of physical systems for feedback control
generally involves a balance between model accuracy and model simplicity [102]. Should

33
a representative mathematical model be difficult to obtain, due to uncertainty or sheer
complexity, conventional techniques prove to be less useful. Also, even though an
accurate model may be produced, the underlying nature of the model may make its
utilization using conventional control design difficult.
Neural networks, hence, have been used for different purposes in the context of
control due to its ability to learn an essential feature of unknown plants by mimic
[43,105,78]. Most of methods existing are based on inverse control. We will therefore
start by considering a neural network controller, specifically, TDNN since this is very
suitable to create both a model and a controller when only input-output measurements are
available.
Principally, the TDNN is an extended multilayer perceptron that allows us to
handle temporal patterns and the problems of time variant signals, i.e., signals that are
scaled and translated over time [40]. The idea that has been followed in the TDNN is
based on the invention of time delays, resulting in giving the individual PEs the ability to
store the history of their input signals. This way, the network as a whole can adapt not
only to a set of patterns, but also to a set of sequences of patterns. An advantage of the
TDNN is the relatively simple mathematical analysis and ability of training by
Backpropagation algorithm. Thus we compare the performance of the proposed control
systems with that of an inverse controller trained through the TDNN.
The algorithm is derived as follows: First we train a TDNN as a model, f , by
letting y k +1 = f ( y k , y k 1 ,L, u k , u k 1 ,L) as shown in Figure 3-2(a). Then a TDNN
controller is designed based on the created TDNN model from which we obtain the
Jacobian

of

the

plant.

The

controller

is

described

by

34
u k = g (u k 1 , u k 2 , L, rk +1 , y k , y k 1 ,L ,W ) letting g be the function implemented by the

controller G ( z ) . Here W is the weights of the TDNN model, F ( z ) . The controller

parameters in the fixed control structure are adapted by an algorithm that ensures that the
desired performance level is maintained and the parameters are updated by back
propagating the error through the model as shown in Figure 3-2(b).
uk

yk

P( z )

Embedding

F ( z)

y k +1

rk +1

rk +1

G( z)

uk

P( z )

yk

Embedding

F ( z)

(a)

(b)

Figure 3-2. Modeling and control scheme using the TDNN: (a) TDNN modeling of a
plant (b) An inverse controller via Backpropagation through (Plant) Model.
3.3 Multiple Inverse Control
Now we discuss the control problem for the local linear model using an inverse
control framework [17]. The central advantage of such a framework is that an inverse
model can be used directly to build a feed-forward controller. When given a model
y k +1 = f i ( y k , y k 1 , u k , u k 1 ) , the control network is brought on line and the control signal
is calculated at each instant of time by setting the output value y k +1 at instance k + 1
equal to the desired value rk +1 as u k = f i 1 ( y k , y k 1 , rk +1 , u k 1 ) while trained off line as in
classical inverse control approach. Thus, for the desired behavior, the controller simply
asks the model to predict the action needed.

35
As stated before, our principal objective is to determine a control input, u k , which
will result in the output, y k +1 , of the plant tracking with sufficient accuracy a specified
sequence, rk +1 . The system identification block has N predictive models denoted by

{ f i }iN=1 , in parallel. Corresponding to each model

f i , a controller g i is designed such that

g i achieves the control objective for f i . Therefore, at every instant one of the models is
selected and the corresponding controller is used to control the actual plant. In order to
control a plant, consider the control problem where the dimension of the input is equal to
that of the output, that is, p = q . From (2.23), because p = q , and under the assumption
that bo is invertible, the control law of an inverse controller for the model, f i o , can be
directly calculated as
1
io ,0

uk = b

dy 1
du 1

rk +1 ai o , j y k j bi o , j u k j

j =0
j =1

(3.1)

Therefore, at time instance k, the control u k can be obtained, if the future target of y k ,

rk +1 , is known. Therefore, the set of local linear models simplifies the control design for a
nonlinear plant. So instead of a global neuro-controller as in other adaptive control
schemes [12,18,31,58], here we can function with a group of linear controllers associated
with each identified model, thus taking care of the system over the whole operating
region.
One advantage of this scheme is its simplicity and fast convergence to get the
desired response. Another advantage is that the dynamic space is decomposed in the
appropriate switching among very simple linear models, which leads to accurate
modeling and controls. On the other hand, creating a set of models by embedded input

36
and output may cause serious problem in the presence of large noise or outliers since the
wrong predictive model due to noise may cause poor control. Hence, the selection of the
right model is as important as creating models and designing controllers.
Once the right local linear model is determined, the corresponding controller is
designed using (3.1). A schematic diagram of the proposed SOM based inverse control
system is shown in Figure 3-3 where the inverse control seeks to model the inverse of the
plant. A set of controllers appears in series with the plant. The command input, rk +1 , is
fed to the controller and provides also the desired response. Hence, when the error is
small the controller transfer function is the inverse of the plant.

uk

yk

Plant P( z )
Embedding

Embedding

rk +1
f11 = g1

f1

f 21 = g 2

f2

f N1 = g N

M
fN

SOM

Figure 3-3. Proposed SOM-based inverse control scheme.

Generally, an adaptive controller that meets the specifications is slow to adapt.
However, our approach models all the operating regions and automatically divides the
operating regions by the number of PEs. So once the current operating region is
determined by the SOM, the corresponding controller is triggered so that the plant tracks
the desired signal. Moreover, even if the wrong PE is assigned in the winning PE due to
noise, a similar dynamic model can be activated since neighboring SOM PEs represent

37
neighboring regions in the dynamic space. Thus, the proposed control system can reach
the set point quickly, and even if the dynamic model is not the most appropriate, there is
an extra flexibility to match the set point with the least amount of error.
3.4 Multiple PID Control
In linear control theory, despite the development of more sophisticated control
strategies, Proportional-Integral-Derivative (PID) controllers have been extensively
studied by researchers and well understood by practitioners, since they are widely used in
practice, and their principle is well understood by engineers [95,92]. It gained its
popularity for its simplicity of having only three parameters. But owing to its simplicity it
has also paid the price of not having an efficient and practical way of determining
optimal gains.
The ideal continuous time PID controller is expressed in Laplace form as follows:
G ( s) = K p +

Ki
+ Kd s
s

(3.2)

where K p is the proportional gain, K i the integral gain, and K d the derivative gain.
Each of the terms works independently of the other1. The standard PID control
configuration is shown in Figure 3-4. The introduction of integral action facilitates the
achievement of equality between the measured value and the desired value, as a constant
error produces an increasing controller output. The derivative action indicates that
changes in the desired value may be anticipated, and thus an appropriate correction may
be added prior to the actual change. Thus, in simplified terms, the PID controller allows
contributions from present, past and future controller inputs.
1

This is not exactly true since the whole thing operates in the context of a closed-loop. However, at any
instant in time, this is true and makes working with the PID controller much easier than other controller
designs.

38

G ( s)
Kp
R( s)

E ( s)
_

Ki s

U (s)

P( s)

Y ( s)

Kd s

Figure 3-4. PID controlled system.

Extensions of the PID control methodology to nonlinear systems, however, are not
trivial. Usually, global nonlinear models are, as always, linearized and the parameters of
the PID controller are scheduled according to the regime. All these difficulties, in fact
stem from the fact that the system model is constructed from a nonlinear dynamical
equation. This difficulty could be eliminated by a piece-wise linear approximation of the
nonlinear dynamics (as one does in gain scheduling). However, gain scheduling methods
are not flexible for model uncertainties and they cannot be scaled up to regimes that are
not described in the initial system identification stage by a linear model. It also has the
problem of either inefficient use of system identification data due to a large number of
arbitrarily selected operating (linearization) points or inaccurate modeling due to the
small number of linearization points [83].
The multiple PID control design method described in this section is rooted in the
principle of using local linear models to construct a globally nonlinear (piecewise linear)
system model that is determined completely from the input-output data collected from the
actual plant [12,30]. In addition to previously designed multiple inverse control schemes,
this section illustrates how the multiple models can be united with the well-known linear
PID controller design techniques to obtain a principled and simple nonlinear PID

39
controller design methodology. The proposed closed loop scheme is illustrated in Figure
3-5 where the SOM determines which linear PID controller contributes to the
instantaneous control input.

G1 ( z )
G2 ( z )

SOM

dk

uk

rk +1

Embedding

P( z )

y k +1

GN ( z)

Figure 3-5. Overall schematic diagram of the nonlinear PID closed loop control
mechanism using multiple controllers.
Once system identification is complete, the design of a globally piece-wise linear
PID control system can be easily accomplished using standard techniques. The literature
has an abundance of PID design methodologies for linear SISO systems including direct
pole-placement techniques and optimal coefficient adjustment according to some criteria
[95,92]. Here pole-placement technique [4] is utilized to design a PID controller for each
linear SISO model and is illustrated briefly in the following.
Assume that the plant is modeled by a set of the SISO, second-order system given
by

y k +1 = f i ( y k , y k 1 , u k , u k 1 )
= a1,i y k + a 2,i y k 1 + b1,i u k + b2,i u k 1

(3.3)

where y k is the output of the plant, and u k is the input to the plant. The controller is
designed by starting with a general PID regulator shown in Figure 3-6 and determining

40
the coefficients of polynomials C i and Di so that the closed-loop system has desired
properties.

R( z + 1)

E (z )

D ( z)
Gi ( z ) = i
Ci ( z )

U (z )

B ( z)
Fi ( z ) = i
Ai ( z )

Y ( z + 1)

Figure 3-6. Block diagram of PID controller for a SISO plant model.
Each models transfer function is given by
Fi ( z ) =

b1,i z 2 + b2,i z
Bi ( z )
= 2
Ai ( z ) z a1,i z a 2,i

(3.4)

It is assumed that the polynomials Ai ( z ) and Bi ( z ) do not have any common factors.
Note that in order to have a stable zero (zero inside the unit circle), b1,i > b2,i must be
true. The controller specifications are expressed in terms of a model that gives the desired
response to command signals. The general control equation is

Ci ( z )U ( z ) = Di ( z )(R( z ) Y ( z ) )

(3.5)

It is assumed that the Ci is monic. To make sure that low-frequency disturbances give
small errors, Ci is chosen as
'

Ci ( z ) = ( z 1) L Ci ( z )

(3.6)

with a suitable selection of L . This puts a ( z 1) term in the denominator of the

controller transfer function, guaranteeing integral control action.
The goal of the controller design is to map the values of a1,i , a2,i , b1,i , and b2,i into
the controller coefficients, d j ,i and ci , subject to the constraint of the model whose

41
characteristic polynomial is z 2 + 1 z + 2 . The transfer function of the closed-loop
system is 1 (1 + zAi ( z )Ci ( z ) / Bi ( z ) Di ( z ) ) ; therefore, its characteristic polynomial is

Ai ( z )Ci ( z ) + z 1Bi ( z ) Di ( z ) . The characteristic polynomial for the model is Am , with the
addition of a second-order observer polynomial term Ao . Equating characteristic
polynomials results in the design equation
'

Ai ( z )( z 1) L Ci ( z ) + z 1 Bi ( z ) Di ( z ) = Ao ( z ) Am ( z )

(3.7)

which can be written as

( z 2 a1,i z a2,i )( z 1)( z + ci ) + z 1 (b1,i z 2 + b2,i z )(d1,i z 2 + d 2,i z + d 3,i ) = z 2 ( z 2 + 1 z + 2 ) (3.8)

This equation is solved for the d j ,i and ci . Then the control law difference equation is

u k = (1 c i )u k 1 + c i u k 2 + d 1,i e k + d 2,i e k 1 + d 3,i e k 2

(3.9)

Given the local linear models as obtained through the use of a SOM and a PID
design technique, the overall closed loop nonlinear PID design reduces to determining the
coefficients of the individual local linear PID controllers using their respective linear
plant model transfer functions. One needs to determine a set of PID coefficients per linear
model. In the competitive SOM approach, the model output depends only on a single
linear model at a given time; therefore, the PID coefficients are set to those values
determined for the instantaneous winning model.

CHAPTER 4
MULTIPLE QUASI-SLIDING MODE CONTROL
Attempts to use traditional control methods, such as inverse control and PID
control, for nonlinear plants will inevitably encounter problems when faced with the
nonlinear nature of these systems. In order to overcome these difficulties in designing
controllers for nonlinear systems, a simplified control method that keeps the advantage of
the conventional approach was proposed in Chapter 3, i.e., the SOM was explored as a
modeling infrastructure, and the controllers were built based on SOM-based multiple
models.
However, these control methods (especially multiple inverse control scheme) may
show poor control performance when existing sensor noise or external disturbances due
to the fact that the perfect control is achieved if the plant and the controller is stable, if
the model is perfect, and if there is no disturbance. Wrong selection of the model caused
by noise can devastate the control mission since the controller is also determined by very
different model from the given task. For this reason, sliding mode control architecture,
which is a very well-known robust controller, is employed in this chapter in order to
obtain sturdiness against noise as well as uncertainties on the model.
4.1 Introduction to Variable Structure Systems
The control of nonlinear systems has been an important research topic and many
approaches have been proposed. While classical control techniques have produced many
highly reliable and effective control systems, great attention has been devoted to the
design of variable structure control systems (VSCS). Variable structure systems (VSS)
42

43
are a special class of nonlinear systems characterized by a discontinuous control action
which changes structure upon reaching a set of switching hyperplanes during the control
process to attain improved overall characteristics in the controlled system. During the
sliding mode the VSCS has invariance properties, yielding motion that is remarkably
good in rejecting certain disturbances and parameter variations [93,9,25,86].
Most of the VSCS proposed in the literature have been developed mainly based on
the state-space model with the assumption that all state variables are measurable or on the
input-output model for a linear system. But in some control problems, we are allowed to
access only the input and the output of the nonlinear plant [72]. In this case, an observer
could be used to estimate the unmeasurable state variables if the state equations are
known. Otherwise, this is not possible. Thus, it is the purpose of this chapter to provide a
new technique to design sliding mode control systems based on input-output models of
the considered discrete-time nonlinear system so that the amount of guesswork1 is
reduced, while attainable performance is increased.
Normally, the design of VSS consists of two parts: First, the sliding surface, which
is usually of lower order than the given process, must be constructed such that the system
performance during sliding mode satisfies the design objectives, in terms of stability,
performance index minimization, linearization of nonlinearities, order reduction, etc.
Second, the switched feedback control is designed such that it satisfies the reaching
condition and thus drives the state trajectory to the sliding surface in finite time and
maintains it there thereafter [93].

In most cases, we need to estimate the unknown parameters, unmodeled dynamics and bounded
disturbances. Also, it should be noted that the VSCS works best when the plant is completely known.

44
4.1.1 Sliding Hyperplane Design
Sliding surfaces can be either linear or nonlinear. The theory of designing linear
switching surfaces for linear dynamic system has been developed in great depth and
completeness. While the design of sliding surfaces for more general nonlinear systems
remains a largely open problem. For simplicity, we focus only on linear switching
surfaces. Moreover, for surface design, it is sufficient to consider only ideal systems, i.e.,
without uncertainties and disturbances. Consider a general system

x& = A( x) + B( x)u

(4.1)

with a sliding surface

S = {x | S ( x) = 0}

where A( x) , B( x) are general nonlinear functions of x , and x n , u m .

The equivalent control is found by recognizing that S& ( x) = 0 is a necessary
condition for the state trajectory to stay on the sliding surface S ( x) = 0 . Therefore,
setting S& ( x) = 0 , i.e.,

S
S
S
S& ( x) = x& =
A( x) +
B( x)u eq = 0
x
x
x

(4.2)

yields the equivalent control

1

S
S
u eq =
B( x)
A( x)
x
x

(4.3)

where (S / x) B( x) is nonsingular. When in sliding mode, the dynamics of the system is

governed by
1

S
S

x& = I B( x)
B( x)
A( x)

x
x

(4.4)

45
For example, if the system (4.1) is linear and described by
x& = Ax + Bu

(4.5)

where A and B are properly dimensioned constant matrices. The switching surface can
be defined as
S ( x) = C T x = 0

(4.6)

i.e., S / x = C T , where C = [c1 , c 2 , L , c m ] is an m n matrix, and then we have

T

u eq = (C T B ) 1 C T Ax

(4.7)

and (4.4) becomes

x& = ( I B(C T B) 1 C T ) Ax = ( A BK ) x

(4.8)

(4.4) and (4.8) describe the behavior of the systems (4.1) and (4.5), respectively, which
are restricted to the switching surface if the initial condition x(t 0 ) satisfies S ( x(t 0 )) = 0 .
For the linear case, the system dynamics is ensured by a suitable choice of the feedback
matrix K = (C T B) 1 C T A . In other words, the choice of the matrix C can be made
without prior knowledge of the form of the control u .
4.1.2 Sliding Mode Control Law Design
Once the sliding surfaces have been selected, attention must be turned to solving
the reachability problem. This involves the selection of a state feedback control function
u : n m which can drive the state x towards the surface and thereafter maintains it
on the surface illustrating in Figure 4-1. In other words, the controlled system must
satisfy the reaching conditions.
The classic sufficient condition for sliding mode to appear is to satisfy the
condition s i s&i < 0, i = 1,L, m and a similar condition proposed by Utkin [93], i.e.,

46
lim s&i < 0 and lim s&i > 0 . These reaching laws result in a VSC where individual

si 0 +

si 0

switching surfaces and their intersection are all sliding surfaces. This reaching is global
but does not guarantee finite reaching time.
x&

si ( x) = 0
x

Reaching phase

Sliding phase

Reaching phase

Figure 4-1. Phase plane plot of a continuous-time second-order variable structure system.
Another commonly used reaching law is proposed by Gao & Hung [27]. The law
directly specifies the dynamics of the switching surface by the differential equation
S& = Q sgn( S ) Rg ( S )

(4.9)

where the gains Q and R are diagonal matrices with positive elements, and
sgn( S ) = [sgn( s1 ),L , sgn( s m )] , g ( S ) = [g 1 ( s1 ),L , g m ( s m )] , where
T

sgn( si ) = {1 if si ( x) > 0, 0 if s i ( x) = 0, 1 if s i ( x) < 0}

(4.10)

and the scalar functions g i satisfy the condition

s i g i ( s i ) > 0 , when si 0

(4.11)

Various choices of Q and R specify different rates for approaching S and yield
different structures in the reaching law.

47
For one way of designing controllers, recall that during sliding mode, one can
compute the equivalent control u eq according to (4.3) or (4.7). However, only using u eq
cannot drive the state towards the sliding surface if the initial conditions of the system are
not on S . One popular design method is to augment the equivalent control with a
discontinuous or switched part, i.e.,
u = u eq + u N

(4.12)

where u eq is a continuous control defined by (4.3), and u N is added to satisfy the

reaching condition which may have different forms. For a controller having the structure
of (4.12), we have

S
S
S& ( x) =
x& =
A( x) + B( x)(u eq + u N )
x
x
S
S
=
A( x) + B( x)u eq +
B( x)u N
x
x
S
B( x)u N
=
x

for simplicity, assume (S / x) B( x) = I , then S& ( x) = u N . Some often used forms of u N

are relay type of control u N = sgn( S ( x)) , linear continuous feedback control
u N = ( S ( x)) , and linear feedback control with switching gains u N = x where
= [ ij ] is an m n matrix and
ij
ij =
ij

if s i ( x) x j > 0
if s i ( x) x j < 0

Parameters ij and ij are chosen to satisfy the desired reaching condition.

48
Another design method of controllers is to employ the reaching law approach
proposed by Gao & Hung [27] and can be directly obtained by computing S& (x) along the
reaching mode trajectory, i.e.,

S
S
[A( x) + B( x)u ] = Q sgn( S ( x)) Rg ( S ( x))
S& ( x) =
x& =
x
x

(4.13)

Hence, we have
T

S
S

u = B( x)
A( x) + Q sgn( S ( x)) + Rg ( S ( x))
x
x

(4.14)

By this approach, the resulting sliding mode is not preassigned but follows the natural
trajectory on a first-reach-first-switch scheme. The switching takes place depending on
the location of the initial state.
4.2 Sliding Mode Control in Sampled-Data Systems
The VSS theory which was originally developed from a continuous time
perspective has been realized that directly applying the theory to discrete-time systems
will confront some unconquerable problems, such as the limited sampling frequency,
sample/hold effects and discretization errors [2]. Since the switching frequency in
sampled-data systems cannot exceed the sampling frequency, a discontinuous control
does not enable generation of motion in an arbitrary manifold in discrete-time systems
[3,9]. This leads to chattering along the designed sliding surface, or even instability in
case of a too large switching gain [2]. Figure 4-2 illustrates that in discrete-time systems,
the state moves around the sliding surface in a zigzag manner at the sampling frequency.
Much research has been done in this field. Among various concepts of discrete-time
sliding mode, Quasi-Sliding Mode (QSM) is reviewed in this section for the sliding mode
controller design aimed at sampled-data systems.

49
State trajectory
k
k+1
k+2

Sliding surface s = 0
Figure 4-2. Discrete-time system response with sliding mode control.
4.2.1 Quasi-Sliding Mode
Many researchers have either addressed the limitations when direct implementation
is done or have proposed designs that take the sampling process into account.
Milosavljevic [57] was among the first researchers to formally state that the sampling
process limits the existence of a true sliding mode. In light of this, the concept of QSM
has been suggested and the conditions for the existence of such mode have been
investigated. Consider a sampled-data SISO system with the predefined sliding surface
x k +1 = Ax k + Bu k

(4.15)

sk = C T xk = 0

The desired state trajectory of a discrete-time VSC system should have the
following features: Firstly, the trajectory moves monotonically towards the switching
manifold and crosses it in finite time starting from any initial point. Secondly, the
trajectory crosses the manifold in succession after it hits the manifold, resulting in a
zigzag motion about the switching surface. Lastly, the trajectory keeps on within a
specified band without increasing the size of each successive zigzagging step. Gao et al.
[28] defined a QSM as the motion of a discrete VSC system satisfying last two features.
In addition they named the specified band, {x | < s ( x) < + } , which contains the QSM
as the Quasi-Sliding Mode Band (QSMB). For

single

input

systems,

the

main

50
approaches for the design of QSM control laws can be categorized into the following two
methods:
Discrete Lyapunov function based approach: Sarpturk et al. [81] noticed that unlike
the case in continuous-time SMC, the switching control in the discrete-time case should
be both upper and lower bounded in an open interval, in order to guarantee the
convergence of sliding mode. Recall that in continuous-time SMC, the control (4.12) is
composed of the equivalent control and a switching control. Converting this control to
discrete-time gives
u k = u k ,eq + u k , N

(4.16)

Hui & Zak [36] observed that if u k , N is a relay control with a constant amplitude,
the relay must be turned off in some neighborhood of the surface, in order to reach the
switching surface, otherwise, the trajectory will chatter around the surface with a chatter
amplitude at least as large as the amplitude of the relay output. The idea of sliding sector
[26,48] was used to solve this problem, i.e., to specify a region in a neighborhood along
the sliding mode, where linear control is used to keep the state inside the region after it
has reached the region. The switching control is applied only when system states are out
of the region. In this case, the derived switching surface is different from the sliding
surface. Based on a discrete Lyapunov function,

Vk =

1 2
Sk
2

the reaching law is given by

S k ( S k +1 S k ) <

1
( S k +1 S k ) 2 for S k 0
2

which ensures Vk +1 < Vk . Furuta [25] proposed a control law of the type

(4.17)

51
u k = u k ,eq + FD x k

(4.18)

where the equivalent control u k ,eq is the solution of

S k = S k +1 S k = 0

(4.19)

and therefore the equivalent control for the system (4.16) is

u k ,eq = (C T B ) 1 C T ( A I ) x k

(4.20)

FD is a discontinuous control law which will be zero inside the sliding sector.

Reaching law based approach: Gao et al. [28] pointed out that the reaching law
(4.17) was incomplete for a satisfactory guarantee of a discrete-time sliding mode, since
it does not ensure that the trajectory moves monotonically towards the switching surface
and the trajectory stays on within a specified band. Thus, they presented an algorithm that
drives the system state to the vicinity of a switching hyperplane in the state space, rather
than to a sector of a different shape. They specified desired properties of the controlled
systems and proposed a reaching law based approach for designing the discrete-time
sliding mode control law. The equivalent form of the reaching law for discrete-time SMC
extended from the continuous-time reaching law (4.13), and for a SISO system is
s k +1 s k = qT sgn( s k ) rTs k , r > 0 , q > 0 , 1 rT > 0

(4.21)

where T > 0 is the sampling period. The state reaches the switching surface at a constant
rate qT and the term rT forces the state to approach the switching surfaces faster
when s k is large. The inequality for T guarantees that starting from any initial state, the
trajectory will move monotonically towards the switching surface and cross it in finite
time. Then the control law for discrete SMC is derived by comparing
s k +1 s k = C T x k +1 C T x k = C T Ax k + C T Bu k C T x k

(4.22)

52
with the reaching law (4.21), which yields,

u k = (C T B) 1 C T Ax k C T x k + qT sgn( s k ) + rTs k

(4.23)

4.2.2 Quasi-Sliding Mode Control Using Multiple Models

Now we discuss the design of the control law for local linear models using the
QSM control framework proposed by Gao et al. [28], where the system states move in a
neighborhood around the sliding surface s k = 0 . The central advantage of the sliding
mode control strategy is that it is an effective robust control strategy for incompletely
modeled or uncertain systems. Thus, the feature of the proposed control scheme is that
the robustness for disturbances can be obtained by the simple control logic based on the
linear model for each region. Another feature of the strategy is that it guarantees
convergence of the system output to a vicinity of the predetermined, fixed plane in finite
time, specified a priori by the designer.
Consider one of the local single input-output models f i of the plant f ,

y k +1 = a1 y k + a 2 y k 1 + L + a m y k m +1 + b1u k + b2 u k 1 + L + bn u k n +1

(4.24)

Equivalently, the input-output model of the plant in (4.24) can be written as the statespace model2
r
r
x k +1 = x k + 1u k + 2 u k 1 + L + n u k n +1

(4.25)

r
where x k = [ y k m +1 , L , y k 1 , y k ]T m is the system state vector which is available for
measurement and and 1 , L , n have the following forms:

In a formal state-space model, past values of the input should be included in the state vector using delay
operators. For simplicity, we include only the system outputs past values in the state vector in this
notation.

53

0
0

= M

0
a m

a m 1

a m2

a m3

0
L 0
M M

L 1
L a1
L

0
0
0
0
0
0



1 = M , 2 = M , L , n = M



0
0
0
bn
b1
b2
Also defining the tracking error vector as
r
r
r
e k +1 = rk +1 x k +1

(4.26)

r
where the desired signal vector is rk +1 = [d k m + 2, L , d k , d k +1 ]T , the switching surface is

defined in the space of the tracking error vector given by

r r
s k = c T ek

(4.27)

r
where c = [c1 , c 2 , L , c m ]T . Then an equivalent control is designed to satisfy the ideal
quasi-sliding mode condition, s k +1 = s k = 0 , by
r
r r
r
r
r
u keq = (c T 1 ) 1 {c T ( rk +1 x k ) c T 2 u k 1 L c T n u k n +1 }

(4.28)

and the closed-loop system response of the ideal quasi-sliding mode substituting (4.28)
into with an equivalent control is given by

r
r
r
r
r
r r
x k +1 = I 1 (c T 1 ) 1 c T x k + 1 (c T 1 ) 1 c T rk +1

(4.29)

r
The system (4.29) can be viewed as a linear system with the input rk +1 and the output
r
x k +1 . To get an insight into the tracking capability of the system, (4.29) can be
represented in terms of the tracking error e k = d k y k by

54

e k +1 =

c m 1
c
c
e k m 2 e k 1 L 1 e k m + 2
cm
cm
cm

(4.30)

Note that by designing the switching surface such that the roots of polynomial

m 1 + (c m 1 / c m ) m 2 + L + (c1 / c m ) are inside of the unit circle, the error vanishes and
thus the condition ensures asymptotic convergence to the desired output. An arbitrary
positive scalar c m also determines the time taken to reach the sliding surface and can be
adjusted to get a faster response.
The reaching law (4.21) always satisfies the reaching condition such that the
discrete VSC system designed using the reaching law approach is always stable with a
stable ideal quasi-sliding mode [28]. Then the control law is derived by comparing

r r
r r
s k +1 s k = c T ek +1 c T e k
with the reaching law (4.21), which yields,

r
u k = c T 1

) [cr
1

r r r
r
rk +1 c T x k c T 2 u k L c T n u k n +1

+ (rT 1) s k + qT sgn(s k )]

(4.31)

Salient feature of the multiple quasi-sliding mode controller is that one can obtain
faster convergence to get the desired response due to multiple control scheme and one
can employ VSCS to control unknown nonlinear plants while gaining indemnity against
noise and parameter variations.
4.3 Analysis of Multiple Quasi-Sliding Mode Control with an Imperfect Sensor
For the SOM-based system identification, one needs to quantify the effect of the
modeling error that will occur due to the quantization of state space induced by the SOM,
and also by the wrong selection of the winning model. Consider model (4.24) in the
presence of modeling error and measurement noise. The predicted output becomes:

55
y k +1 = a1 ( y k + k ) + a 2 ( y k 1 + k 1 ) + L + a m ( y k m +1 + k m +1 )
+ b1u k + b2 u k 1 + L + bn u k n +1

(4.31)

where a wrong local model a1 , a 2 , L , a m , b1 , b2 , L , bn is triggered by the SOM due to

the noisy output measurement y k + k . Then, when s k = 0 , the overall tracking error
response with an equivalent control is given by
c m 1
c
c
e k m 2 e k 1 L 1 e k m + 2
cm
cm
cm
+ (a1 a1 )( y k + k ) + L + (a m a m )( y k m +1 + k m +1 )
+ (b b )u + L + (b b )u

e k +1 =

k 1

(4.32)

k n +1

For simplicity, consider the error dynamics (4.32) when m = 2 . Defining model
r
r
parameter error k = [ a1 a1 , a 2 a 2 , b1 b1 ]T and noise n k = [ k , k 1 , k 2 ]T

r r r
e k +1 = c e k + Tk ( z k + n k )

(4.33)

r
where z k = [ y k , y k 1 , u k 1 ]T , c = c1 / c 2 , and c is chosen as c < 1 . We assume that
r
E[ k2 ] = and k

< wk w k

where the norm wk w k is the Euclidean distance

between the reference vector of the correct PE and that of the neighboring PE selected by
the perturbed output measurements3. Also, it is assumed that the noise k is zero-mean
and white. We have the following recursive formula for the tracking mean squared error.
r
r r
r r
r
r
r r
E[e k2+1 ] = E[c 2 e k2 + 2c e k Tk ( z k + n k ) + Tk ( z k + n k )( z k + n k ) T k ]
r r r r
r
r r r
= c 2 E[e k2 ] + Tk z k z kT k + Tk E[n k n kT ] k

(4.34)

When we take the norm on each side in (4.34), the norm of the tracking error power is
represented by
3

The first assumption states that measurement noise has finite power. The second assumption means model
parameter error is bounded by the distance in the state space.

56
r
E[e k2+1 ] c 2 E[e k2 ] + k

r
zk

r
+ k

(4.35)

where the Cauchy inequality is used on the second term on the right hand side. As
k , using the earlier assumptions on noise and model error bounds, the following

bound on the steady-state tracking error power is obtained:

E[e ] <
2

wk w k
(1 c )
2

[ zr

2
k

(4.36)

Note that the difference between the true model (winning PE) and the neighboring
model (wrong PE) assigned by noisy input, wk w k

, is typically small, since

neighboring SOM PEs represent neighboring regions in the dynamic space. Also, it
should be noted that the error can still be very large if we choose c as close as 1. In
contrast, by choosing c as small as possible, the closed-loop system may have very fast
transient response, possibly too large unexpected overshoot. Thus we should be careful
for determining c so as not to have large error. This problem will be discussed later in
simulation results. If c is set to small enough it then follows that the error by choosing
appropriate design parameters mentioned above will be bounded for a given modeling
uncertainty and measurement noise bounded by . Moreover, this shows that the
switching scheme does not create an issue to be considered in order to guarantee BIBO
stability of the overall system.

CHAPTER 5
CASE STUDIES
To examine the effectiveness of the proposed multiple controller design
methodologies, the chaotic systems with the input term, discrete-time systems, and flight
vehicles have been considered assuming the following:
Assumption 1: The only state available for measurements is y k = x1,k .
Assumption 2: The nonlinear function f is completely unknown.
By assuming that the function f is unknown, we confront a worst case (least prior
knowledge) control design. If an estimate f of f is available, it can be included in the
control design. As can be expected, the better the estimate f , the better the performance
of the resulting controller. Our objective is to design multiple controllers for unknown
nonlinear plants that guarantees global stability and forces the output, y k , to
asymptotically track the desired signal, i.e., y k rk 0 , as k without any a priori
knowledge of the plant.
5.1 Controlled Chaotic Systems
Chaotic systems exhibit irregular, complex, and unpredictable behavior that exists
in many industrial systems. Recently considerable effort has been in the focus of attention
in the nonlinear dynamics literature since removing chaos can improve system
performance, avoid fatigue failure of the system, and lead to a predictable system
behavior. In the literature, several design techniques have been applied for the control
and synchronization of a variety of chaotic systems. A generalized synchronization of
57

58
chaos via linear transformation [106], adaptive control and synchronization of chaotic
systems using Lyapunov theory [103,107], nonadaptive and adaptive control systems
based on backstepping design techniques [31,32,99,100], and an adaptive variable
structure control system for the tracking of periodic orbits [108] have been considered, to
name a few. The contribution of this work lies in the design of multiple control system
for the control of chaos based on the theory of the inverse control, PID control, and
sliding mode control [14] assuming that the chaotic system is unknown. Given an
unknown chaotic system, the goal is to force it to set points or a stable trajectory.
5.1.1 The Lorenz System
The Lorenz model is used for fluid conviction that describes some feature of the
atmospheric dynamic [73]. The controlled model is given by
x&1 = (x 2 x1 )
x& 2 = x 2 x1 x3 + x1 + u
x& 3 = x1 x 2 x3

(5.1)

where x1 , x 2 , and x3 represent measurements of fluid velocity and horizontal and vertical
temperature variations, respectively. The Lorenz system can exhibit quite complex
dynamics depending on the parameter values. For 0 < < 1 , the origin is a stable
equilibrium point. For 1 < < * ( , ) := ( + + 3) ( 1) , the system has two

stable equilibrium points ( 1) , ( 1) , ( 1) and one unstable equilibrium

point at the origin. For * ( , ) < , all three equilibrium points become unstable and
the system trajectory have chaotic behavior [109]. As in the commonly studied case, we
select = 10 , = 8 / 3 which leads to * ( , ) = 24.74 . Thus we will consider the
system with = 10 , = 8 / 3 , and = 28 , which produces the well-known butterfly

59
chaotic dynamics without any control, i.e., u = 0 , as shown in Figure 5-1. Here, the
control force, which is uniformly distributed in u 100 , is added into the second
equation of (5.1).

Figure 5-1. The uncontrolled Lorenz system: phase-space trajectory and time-series.
After solving the set (5.1) with forth-order Runge-Kutta method with integral step
0.05, 7000 samples were generated for analysis. In order to construct a set of modelbased controller, first, we design a model as
y k +1 = f ( y k , y k 1 , u k , u k 1 , u k 2 )

(5.2)

based on the Lipschitz index [35] as shown in Table 5.1.

Table 5-1. Lipschitz index of the controlled Lorenz system for determining an
embedding dimension.
Number of inputs

Number
of
outputs

1
2
3
4
5

1
1090.9
10.5
5.1
3.2
3.0

2
49.7
2.9
2.7
2.3
1.8

3
5.7
2.4
2.0
1.9
1.4

4
3.3
2.1
1.5
1.4
1.4

5
2.1
1.6
1.3
1.2
1.2

60

Figure 5-2. Generalization error v.s. Number of PEs (left) Learning curve (right).
The SOM was trained with the output history vector y ,k = [ y k , y k 1 ]T over
L = 6000 samples with the time decaying parameters, k = 0.1 /(1 + 0.003k )

and

k = ( N / 2) /(1 + 0.003k ) in (2.14), and then local linear models were built from y,k
as well as u ,k = [u k , u k 1 , u k 2 ]T for each PE. A newly generated sequence of M = 1000
samples was applied to multiple models for performance test1. Then the dimension of the
square SOM was determined as 88 based on the generalization error as shown in Figure
5-2 since the error does not decrease much after

N = 8 . Also, with N = 8 8 , the

learning curve2 for 4000 epochs is as shown in Figure 5-2. This curve reflects the overall
closeness of the winning PE to the input samples during the training process and it
becomes approximately constant at the end of training.

Identification performance was evaluated by NRMSE (Normalized root mean squared error):

2
= 1 / max( r ) 1 M kM ( rk +1 y k +1 )
2

The learning curve is defined as the RMS (Root mean squared) distortion between the input and the

winning PE: =

1 L kL | y ,k wi o ,k |

61
The identification result is shown in Figure 5-3 where the dashed line is the output
of the controlled Lorenz system, and the solid line is the output from the multiple models.
As we can see, the multiple models are a very good approximation of the controlled
Lorenz system. In addition, plant modeling performance with 64 multiple models was
compared with a single linear model, ARX, with the same number of inputs used in local
modeling.

Figure 5-3. Identification of the controlled Lorenz system by multiple models.

It was also compared with those by means of a conventional TDNN, which was
trained by the backpropagation algorithm with the constant learning rate of 0.001 for
3000 iterations on the same number of inputs and outputs as in local linear modeling,
adopted as a global nonlinear model and listed in Table 5.2. The best result with the
proposed method was a NRMSE of 0.0131 while with the TDNN3 obtained a NRMSE of

The number of PEs in the hidden layer of the TDNN is chosen as 10 by 20 Monte-Carlo simulations
varying the size of the hidden layer.

62
0.0205. On the other hand, a single linear ARX model produced much higher NRMSE,
0.0486, than others did. This result shows that the proposed multiple linear modeling
scheme outperforms both the linear and the nonlinear global modeling paradigm.
Table 5-2. Comparison of modeling performance for the controlled Lorenz system.
Methodology

NRMSE

ARX (1)

4.8e-2

Multiple ARX (64)

1.3e-2

TDNN (5:10:1)

2.1e-2

Next, we tested the proposed control scheme using multiple models. First, we let
the system converge to the origin, which is one of the equilibria of the Lorenz system,
starting from the initial state [ x1 , x 2 , x3 ] = [10,10,10]T by the Multiple Inverse Controllers
(MIC); the controller is activated at 5 sec and the results are shown in Figure 5-4(a)
where the closed loop system response is seen to converge to the origin fairly well. This
is relatively easy to control since the affine system (5.1) of zero dynamics is
asymptotically stable at the equilibrium point. Also the figure shows the behavior of the
control input. Second, we forced the controlled Lorenz system to a set point, y d = 8
which is not one of the equilibria. This is a more complicated task than steering the state
of the system to the origin. The regulation results with the proposed MIC are shown in
Figure 5-4(b) where we observe that the states x1 and x 2 asymptotically regulate to
x1 = 8 and x 2 = 8 , respectively, and the state x3 remains bounded. They illustrate that the

conditions x 2 = x1 and x3 = ( x1 ) 2 / described in [108] are satisfied even in the case

under consideration.

63

(a)

(b)
Figure 5-4. Tracking a fixed point reference signal by MIC: (a) y d = 0 (b) y d = 8 .
Afterward, the MIC was investigated as increasing the number of controller and
compared with a single linear inverse controller built by a ARX model as well as a
nonlinear TDNN controller, which is a global controller trained through the TDNN
model, for the same task. The optimal number of PEs in the hidden layer of the TDNN
controller was selected as 40. Figure 5-5 illustrates how the number of controllers effects
control performance. As expected, a single inverse controller based on a ARX model
showed the worst performance with regard to settling time as well as steady-state error
even though it reached the set-point the first time. As seen in the figure, the faster the
rising time, the larger the overshoot, which is an unwanted factor in most cases, is

64
generated. A smaller overshoot, faster settling time, and less steady-state errors were
obtained by using a larger number of inverse controllers, such as the 16-MIC, which
showed a very similar performance to the TDNN controller. Moreover, the response
using the 64-MIC demonstrated very small steady-state errors. The 144-MIC, however,
reached the set-point much slower than the others. It also took much longer to settle to
the desired point. Too many divisions of state space may cause poor control performance
(frequent switching among controllers) so that the controller may not be capable of
following fast-changing trajectory.

Figure 5-5. Comparison of control performance varying the number of inverse controllers
based on multiple models.
Finally, the proposed multiple control schemes, especially MIC, were compared
with a linear controller and a nonlinear controller. The PID controller coefficients were
determined to bring the poles of the closed-loop response from the plant output to the
desired output to 0, 0, 0.55+i0.31, and 0.55-i0.31. The QSM controller was designed by

65

r
choosing the switching surface with c = [1,3]T and qT = 0.8, rT = 0.001 . Depending on
the selection of the coefficients for PID and QSM controllers, one obtains the
performance desired. This issue will be discussed in later sections. The step performance
of the designed multiple model based control scheme and the TDNN controller in closedloop operation with the controlled Lorenz system is provided for the two case studies,
y d = 0 and y d = 8 , in Figure 5-6.

Figure 5-6. Comparison of tracking performance by multiple model based controllers

(MIC, MPIDC, MQSMC) and global inverse controllers (IC-ARX,
TDNNC).
We can observe from these results that the overshoots of the Inverse Controller based on
the ARX model (IC-ARX) are much larger than others, which also results in a much
longer settling time. Moreover, even if IC-ARX and TDNNC demonstrated very short
reaching time, they exhibited relatively larger steady-state errors than the others. On the
contrary, the multiple inverse controller showed a much faster reaching time, which is
close to that by IC-ARX and TDNNC, with small steady-state errors when compared
with MPIDC and MQSMC. These comparisons of IC-ARX, TDNNC, and MIC seem
reasonable since they all have an inverse control framework. From these results, it can be
easily inferred that the proposed control strategy guarantees the convergence of the

66
system gaining a fast response to the desired set-point, even though the plant with the
highly nonlinear characteristics is not known a priori to the controller.
5.1.2 The Duffing Oscillator
Another system considered corresponds with the control of a Duffing oscillator,
which displays chaotic behavior, described by
x&1 = x 2
x& 2 = u p1 x 2 p 2 x1 p3 x13 + p 4 cos t

(5.3)

where is a constant frequency parameter, p1 , p 2 , p3 and p 4 are constant parameters

[10,32,34,35].

Figure 5-7. The uncontrolled Duffing oscillator: phase-space trajectory and time-series.
We assume that the controlled Duffing oscillator is originally ( u = 0 ) in the chaotic state,
shown in Figure 5-7, with parameters = 1.8 , [ p1 , p 2 , p3 , p 4 ] = [0.4, 1.1, 1.0, 1.8]T
[32]. In the simulations, the Duffing system was considered as unknown, which only
generated time-series data excited by uniformly distributed control input, u 5 , via the
fourth-order Runge-Kutta scheme with a fixed time step of 0.2. The embedding
dimension for model construction was chosen as d u = 1 and d y = 1 based on the Lipshitz
index in Figure 5-8. Thus the model is designed as y k +1 = f ( y k , y k 1 , u k , u k 1 ) . Another

67
parameter to be selected is the size of the SOM which is trained with ky = [ y k , y k 1 ]T .
We chose the number of PEs in the SOM as 88 by the performance varying the size of
the map shown in Figure 5-8.

Figure 5-8. Lipschitz index (left) for the determination of optimal number of inputs and
outputs and Generalization error v.s. Number of PEs (right).
Table 5-3. Comparison of modeling performance for the controlled Duffing oscillator.
Methodology

NRMSE

ARX (1)

3.4e-2

Multiple ARX (64)

1.3e-2

TDNN (4:14:1)

1.1e-2

The identification result with 88 square map and its performance comparison with
one of global model, TDNN4, is presented in Figure 5-9. In addition, another comparison
with a single linear model is listed Table 5-3. We observe that the proposed multiple
modeling strategy is a little worse than a TDNN, even if it demonstrated much better
modeling performance than a single ARX model. The proposed modeling method,

A TDNN with 14 PEs in the hidden layer demonstrated best performance in modeling the Duffing
oscillator.

68
however, has a much simpler structure for modeling chaotic systems than a global
modeling paradigm.

Figure 5-9. Identification of the controlled Duffing oscillator by TDNN (left) and
multiple-models (right).
Next, we performed simulations for 3 different control tasks.
1) Control of a Duffing oscillator: For this task, we target the oscillator to follow an
arbitrary trajectory generated by a random control input bounded by 5.
2) Synchronization of two Duffing oscillators: This is to show the proposed control
scheme is able to synchronize oscillators effectively in spite of model mismatch; the
parameters for the slave oscillator are p1 = 0.4 , p 4 = 1.8 , and = 1.8 , whereas the
parameters for the master oscillator are p1 = 0.41 , p 4 = 2 , and = 1.9 .
3) Synchronization of two strictly different second order oscillators: In this case,
the slave oscillator is a Duffing one. For this oscillator, the parameters were taken as
p1 = 0.4 , p 4 = 1.8 , and = 1.8 . The master oscillator is a van der Pol one which is
2
3
described by x&1 = x 2 , x& 2 = 0.1( x1 1) x 2 x1 + 1.75 cos* (0.667t ) .

For these missions, the controllers were designed as follows: The optimal number
of PEs in the hidden layer of the TDNNC was selected as 30 by 20 Monte-Carlo

69
simulations. The MQSMC were built by choosing the sliding hyperplane with
r
c = [1,15]T and qT = 0.8, rT = 0.001 . The MPIDC were designed for placing the poles
of the closed-loop response at 0.25 i 0.25 which demonstrated the fastest convergence
to the desired trajectory in Figure 5-10.

Figure 5-10.

(a)

(b)

(c)

(d)

Performance comparison on trajectory tracking by TDNNC, PID-ARX,

and MPIDC when the poles of the closed-loop response are place at (a) 0.9
(b) 0.5 0.5i (c) 0.25 0.25i (d) 0.05 0.05i.

From these figures, it should be noted that the MPIDC produces shorter settling time than
a single PIDC does regardless where the poles are placed. In addition, as the poles are
getting closer to the unit circle or to the origin the convergence time is getting longer.
Hence, the poles of the error dynamics should be chosen carefully.

70
We compared the performance of the controllers using 50sec-long oscillatory
trajectory regarding settling time and NRMS of steady-state error (NRMS-SSE) in Table
5-4: the settling time was selected when the tracking error was bounded in 0.1 and the
NRMS-SSE was calculated using the tracking error from 30sec to 50sec. As seen in the
table, the proposed control strategies outperformed a global controller, which is generally
utilized for unknown system control, in both fast response and accuracy. Especially,
while the TDNNC demonstrated some difficulty in following the desired path generated
from a different dynamics, the van der Pol oscillator, the multiple controllers
accomplished the mission relatively well.
Table 5-4. Comparison of tracking performance for 3 different control task : Settling
time and NRMS-SSE.
Methodology
TDNNC
MIC
MPIDC
MQSMC

Task 1

Task 2

Task 3

6.0sec
2.8e-2
5.0sec
2.3e-2
3.6sec
1.9e-2
4.2sec
2.7e-2

3.6sec
3.2e-2
3.6sec
2.2e-2
1.6sec
1.9e-2
2.4sec
2.6e-2

10.6sec
6.8e-2
7.8sec
2.7e-2
6.2sec
1.8e-2
5.2sec
3.2e-2

Moreover, the MPIDC among the proposed multiple control methodologies showed a
significant reduction of the transient time as well as accuracy improvement in most
missions. Figure 5-11 shows the synchronization of the master and slave oscillator
signals by the TDNNC and MPIDC where the controller is activated at t = 8 sec . It is
seen from the results again that the transient time is shortened without resulting in
overshoots using the MPIDC.

71

(a)

(b)

(c)
Figure 5-11.

Control performance by TDNNC (left) and MPIDC (right): The dotted

line is the desired trajectory and the solid line is the output of the
oscillator. (a) Tracking an oscillatory reference signal (b) Synchronization
of two Duffing oscillators (c) Synchronization of a Duffing oscillator and
a van der Pol oscillator.

72
5.2 Nonlinear Discrete-Time Systems
We have seen how effectively the PID and the inverse controller can be
implemented using multiple models. In this section, nonlinear discrete time systems in a
noisy environment are considered focusing more on multiple model-based control with
sliding mode.
5.2.1 A First-order Plant
Consider the following nonlinear discrete-time plant [5]
x1,k +1 = x 2,k
x 2,k +1 =

3 x1,k

+ x 2,k + u k
16 1 + ( x 2,k ) 2

(5.4)

y k = x 2,k + z k

where u k is the input and z k is an external disturbance. In (5.4), we considered a SISO

model, assuming that only the output y k is available for measurement. The output timeseries was created by exciting an input signal that is uniformly distributed in u 0.5 .

(a)
Figure 5-12.

(b)

Parameter selection to design multiple models: (a) Lipschitz index for

determining the embedding dimension (b) Identification performance v.s.
network dimension on independently generated test data for choosing the
size of a map.

73
The model was assumed to be a second order in input and output based on Figure
5-12(a). After quantization of the embedded output space y , a set of models was built
with the input-output data samples for each PE. For testing, 400 independently generated
data samples were used changing the size of the map. The best size of the map was
determined as 88 since the performance did not improve much after 64 PEs (see Figure
5-12(b)). Thus plant identification with 64 multiple models (88) was tested in the
absence of sensor noise as well as in the presence of sensor noise with the plant input
signal being uniformly distributed. The result of system identification in the absence of
sensor noise by multiple models is shown in Figure 5-13(a). As we can see, the models
provide a very good approximation of the plant visually based on the error signals.

(a)

(b)

Figure 5-13. Modeling performance using 64 multiple models for a nonlinear first order
plant: (a) System identification of the nonlinear plant in the absence of
disturbance by the proposed multiple models. (b) Comparison of robustness
against noise between TDNN and multiple models.
Also, the proposed multiple modeling scheme was compared with a conventional
TDNN that was trained through the backpropagation algorithm with 5000 samples and a
constant learning rate, 0.005. The modeling result with 64 multiple models was a NRMSE

74
of around 6.5e-4 while with the TDNN5 one obtained a NRMSE of about 6.3e-3. The
performance when the testing data are perturbed by noise is shown in Figure 5-13(b)
where we observe that the SOM is more likely to select the wrong model as the noise
level is increased. Even though the multiple models presented more accuracy in
identification than a global model in the absence of noise, it should be pointed out that
the multiple modeling strategy does not have more noise-immunity than global models at
certain noise level. However, the proposed method is more robust than a global model up
to certain noise level.

Figure 5-14.

Responses for parameter selection to design QSMC by varying (a) rT and

(b) qT .

Based on the 64 multiple models and the TDNN model identified, we designed
multiple controllers and a TDNN controller, respectively. The TDNN controller was
trained by back-propagating an error through the TDNN model taught by 20 hidden PEs.
The number of hidden PEs in the controller was chosen as 40. The MPIDC was designed
for placing the poles of the closed-loop response at 0.1 i 0.3 . The MQSMC was built by

r
choosing the sliding hyperplane with c = [1,2]T and qT = 0.8, rT = 0.01 comparing the
5

The number of PEs in the hidden layer of the TDNN is chosen as 20 by 20 Monte-Carlo simulations
varying the size of the hidden layer.

75
parameters as in Figure 5-14. In the figures, we observe that the larger the value of rT is
the faster the plant reaches to the reference signal. In addition, the smaller the value of
qT is the smaller the sliding mode band is.
First, the performance of the controllers was tested on square-wave [1,-1.5, 1, 0]
tracking in the absence of sensor noise and the results are presented in Figure 5-15 where
we can see that all controllers showed very good performance on tracking the reference
signal. Specifically, the MPIDC is the most accurate controller in spite of very slow
convergence, and the MIC is the fastest one even though it shows a little steady-state
error. In contrast, the MQSMC demonstrated the worst performance among the 4
controllers regarding rising time and steady-state error, but it shows its superiority in the
presence of noise later.

Figure 5-15.

Comparison of tracking performance using a global controller and

multiple model based controllers in the absence of sensor noise. The figure
(right) is an enlargement of the figure (left) between 34 and 52 iterations.

Figure 5-16 shows the plant responses of the closed loop control system using the
MQSMC. The trajectories are seen to converge to the desired values of [-1.5 -1.0 -0.5 0.5
1.0 1.5]. The figure also shows control input, the sliding surface, and the winner activities
switched automatically by the SOM. It can be easily seen that the proposed MQSMC

76
scheme guarantees the convergence of the system to the quasi-sliding-mode band around

r r
the sliding hyperplane c T e k = 0 .

Figure 5-16.

Performance of square-wave tracking in the absence of noise by the

MQSMC.

Next, the robustness of the proposed control scheme was compared with that of a
global controller using TDNN. The standard deviation of the error between the plant
output and the desired output versus the standard deviation of the noise is shown in
Figure 5-17. It is evident that the MQSMC performs best in terms of insensitivity to
disturbances. The MIC structure showed the best performance only in the noise-free
environment. It should be noted that the MIC and the MPIDC began to become less
robust than the TDNNC at the point where the standard deviation of the noise is over
0.04. From this examination we can conclude that the MIC and the MPIDC can be robust

77
against noise up to certain level. However, wrong selection of the winner, as the amount
of noise is increased, can be devastating for the controller that is designed based on the
predicted model.

Standard deviation of tracking error

TDNNC

M IC

M PIDC

M QSM C

0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
0.000

0.003 0.011 0.023 0.046

0.070 0.093 0.115

Standard deviation of noise

Figure 5-17.

Comparison of performance against noise among TDNNC, MIC, MPIDC

and MQSMC.

Furthermore we tested the closed loop system for tracking a sinusoidal and an
arbitrary desired output. Once again, the multiple controller networks perfectly track the
desired command except for a transient time of a few time steps, as shown in Figure 5-18,
even if the measurement is corrupted by zero-mean random noise with 20dB of SNR.
Overall, we conclude that the proposed MQSMC approach is the most robust
design technique among the four methods considered. This is evident from Figure 5-17,
where we observe that on average the tracking error of the MQSMC increases at a lower
rate than that of the MIC, the MPIDC, and the TDNNC.

78

(a)
Figure 5-18.

(b)

Sinusoidal and arbitrary signal tracking by the MQSMC:(a) in the absence

of sensor noise (b) in the presence of sensor noise, SNR = 20dB .

5.2.2 A Laboratory-scale Liquid-level Plant

A liquid-level system described by the following second-order equation has been
considered:
y k +1 = 0.9722 y k + 0.3578u k 0.1295u k 1 0.3103 y k u k
0.04228 y k21 + 0.1663 y k 1u k 1 0.03259 y k2 y k 1
0.03513 y k2 u k 1 + 0.3084 y k y k 1u k 1

(5.5)

+ 0.1087 y k 1u k u k 1
This model is obtained through identification of a laboratory-scale liquid-level system
[80,1]. In [80], this model has been used to illustrate theoretical developments for direct
adaptive control. 10,000 data samples were generated for analysis using the control

79
effort, u k 1 , which has uniform distribution. Given the prior information concerning
the order of the plant by Lipschitz index in Table 5-5, a second order input-output model
described by the following equation was chosen to identify the plant:
y k +1 = f ( y k , y k 1 , u k , u k 1 )

(5.6)

Table 5-5. Lipschitz index of a laboratory-scale liquid-level plant for determining an

embedding dimension.
Number of inputs

Number
of
outputs

1
2
3
4
5

(a)

1
262.34
13.49
1.46
1.19
1.19

2
10.86
1.32
1.16
1.15
1.15

3
1.80
1.18
1.13
1.10
1.09

4
1.64
1.14
1.13
1.10
1.09

5
1.46
1.13
1.12
1.09
1.09

(b)

Figure 5-19. Modeling performance using multiple models for a liquid-level plant: (a)
Identification performance v.s. network dimension for choosing the size of a
map of the liquid-level plant (b) Identification of a liquid-level plant using
1212 multiple models.
The embedded output vector, y,k , was used for SOM training over 6,000 epochs,
and then not only y,k but the embedded input vector, u,k , were exploited to create
local linear models. In order to test the performance of the proposed modeling scheme,

80
we applied 400 newly generated data samples to multiple models. Figure 5-19 depicts
identification performance depending on the size of the network and how well the
multiple models approximate the liquid-level plant using 144 models.
In addition, we compared the identification performance using multiple models
with that by a TDNN. The number of inputs and outputs to the network were the same as
in local linear modeling. A single hidden layer of 30 PEs was large enough for good
identification performance by 20 Monte-Carlo simulations varying the size of the hidden
layer. As seen in Table 5-6, the proposed multiple models slightly outperformed a
nonlinear model in a liquid-level plant identification. Also the table clearly shows the
benefit of using multiple models when comparing with the performance by a single ARX
model.
Table 5-6. Comparison of modeling performance for the liquid-level plant.
Methodology

NRMSE

ARX (1)

29.2e-3

Multiple Models (1212)

3.5e-3

TDNN (4:30:1)

4.3e-3

A typical open-loop input-output characteristic of the plant is shown in Figure 5-20.

The large variations in the steady-state gain and time constant with the operating point is
also clearly visible from this figure. The performance of the proposed multiple model
based controllers were compared to that of a nonlinear TDNN controller designed
through the previously identified TDNN model. The controllers for this plant were
designed as follows: First, the optimal number of PEs in the hidden layer of the TDNN
controller was chosen as 50. Second, the PID controller was designed in order for the
closed loop response poles to be located in 0.7702 and 0.1298. Third, the sliding

81

r
hyperplane of the QSM controller was set to c = [1, 1.66]T based on Figure 5-21 which
illustrates the effect of the parameter selection for the sliding surface design under
different noise levels.

Figure 5-20. Typical input-output characteristic of the second-order liquid-level plant.

Figure 5-21.

Square-wave tracking performance of the liquid-level plant varying the

sliding surface and the noise level by the MQSMC.

82
In most cases, as we placed the poles ( c1 / c 2 ) closer to the origin inside the unit circle,
the controller showed better tracking performance. For instance, from the plot, we can
say that the pole should be chosen as between 0.5 and 0.6 to have the robustness against
noise whose level is 25dB of SNR since the error changes too slowly above 0.8 and very
fast below 0.5. Thus the switching surface was chosen as s k = [1,1.66][ek 1 , ek ]T in
order to get small error and short enough transient time as well.

Figure 5-22.

(a)

(b)

(c)

(d)

Control of the liquid-level plant by the MQSMC varying the number of

controllers: (a) M = 1 (b) M = 16 (c) M = 36 (d) M = 144.

The control systems were tested for a typical square wave set-point of amplitude
0.5. Figure 5-22 compares the tracking performance of the reference signal, which is the
output of a first-order reference model with transfer function Gd ( z ) = 0.2 z /( z 0.8) and

83
driven by the set-point, by the MQSMC varying the size of the map. The figure illustrates
how the control performance is affected by the number of QSM controllers. As the
number of controllers is increased the plant reaches the set-point with smaller steady-state
errors. Even if there seems, however, no big difference between Figure 5-22(c) and
Figure 5-22(d) the larger number of controllers might be working better in a noisy
environment. This will be discussed again later in this section. Additionally, a
comparison of this figure with Figure 5-20 displays the open-loop response and
establishes the efficacy of the proposed control scheme.
Table 5-7. Comparison of control performance for the liquid-level plant in noise-free
environment.
Methodology

NRMSE

TDNNC (4:50:1)

5.7e-3

MIC (1212)

1.4e-3

MPIDC (1212)

10.9e-3

MQSMC (1212)

5.0e-3

Table 5-8. Comparison of control performance for the liquid-level plant in the presence
of sensor noise: standard deviation of noise is 4.5e-2.
Methodology

Standard deviation
of error

TDNNC (4:50:1)

3.8e-2

MIC (1212)

4.2e-2

MPIDC (1212)

10.2e-2

MQSMC (1212)

1.9e-2

Table 5-7 compares the tracking performance among controllers where MIC and
MQSMC outperformed TDNNC. However, even though MPIDC showed the worst
performance it showed the smallest steady-state errors among them. It only demonstrated

84
large errors in the transient phase since it was not fast enough to track the trajectory. In
addition, the controllers were tested on the same tracking trajectory regulating
disturbances whose standard deviation is 4.5e-2. The results are listed in Table 5-8.

Figure 5-23.

(a)

(b)

(c)

(d)

Control of the liquid level system with measurement noise by the

MQSMC with (a) M = 1 (b) M = 16 (c) M = 144 and (d) the TDNNC.

As in the previous result for controlling the first order plant, the MQSMC showed
the best performance in robustness against the measurement noise, and the MPIDC
showed the worst performance again due to the slow convergence. Figure 5-23 compares
the response among a nonlinear controller and the MQSMCs varying the number of the
controllers under a noisy environment. As we can see, the proposed MQSMC is capable
of rejecting measurement noise and following the trajectory with very little variations.

85
Also the MQSMC with M = 144 showed a more stable aspect than the smaller structure.
While a nonlinear controller, the TDNNC, showed relatively large chatters on the
trajectory even if it tracks the path, it also spent much more control effort to do the same
mission than the MQSMC did.

Open-loop Control

MPIDC

TDNNC
Measurement noise
MIC

MQSMC

Tracking an oscillatory reference signal of the liquid-level plant by the

TDNNC (left) and the MQSMC (right) in the presence of sensor noise.

Perfect Cotnrol

Figure 5-24.

Increasing closed-loop variance

Figure 5-25. Performance assessment on a trajectory tracking under noisy environment.

Finally, we evaluated the tracking performance of an oscillatory reference signal by
the TDNNC and the MQSMC depicted in Figure 5-24. The reference signal to be
followed was 0.2(0.8 sin(0.15k + / 2) + 1.2 cos(0.1k )) . We observed that the MQSMC

tracked the trajectory more smoothly than the TDNNC did. In summary, the proposed

86
MQSMC is the best possible multiple model based controller in a noisy environment for
both trajectory tracking and set-point tracking as illustrated in Figure 5-25. On the other
hand, the MPIDC is not suitable for a trajectory tracking problem that requires fast
transient response coping with measurement noise even if it is the best in set-point
tracking in the absence of noise.
5.3 Flight Vehicles
Flight vehicles, such as missiles and aircraft, are very complex systems that are
typically non-minimum phase and have aerodynamic coefficients which vary over a wide
dynamic range due to large Mach-altitude fluctuations [43,47]. Control of highperformance low-cost UAV especially involves the problems of incomplete
measurements, external disturbances and modeling uncertainties. Nevertheless, the
autopilot for these vehicles is often required to achieve very stringent performance
objectives. Thus the proposed control algorithms are applied to missiles and UAV to
show the effectiveness in this section.
5.3.1 Missile Dynamics
We consider a benchmark missile model used widely in the literature [50,104].
This benchmark model can be formulated as
& g1
q& = g
2

1 g 3
+

0 q g 4

and
g1 = [0.0001 2 0.0112 0.2010(2 M / 3)]M cos( )
g 2 = [0.0152 2 1.3765 + 3.6001(7 + 8M / 3)]M 2
g 3 = 0.0403M cos( )
g 4 = 14.542M 2

(5.7)

87
where is the angle of attack (deg), q is the pitch rate (deg/s), M is the Mach number,
and is the tail deflection angle (deg). We considered an operation range of M = 3 and

[8, 8] deg. The model was formulated as y k +1 = f ( y k , y k 1 , u k , u k 1 ) where yk is

the Angle Of Attack (AOA) to be controlled.

(a)

(b)

Figure 5-26. Modeling performance using multiple models for the missile dynamics: (a)
Identification performance varying the network dimension (b) Identification
of a missile dynamics by 1818 multiple models.
The SOM was trained with the vector [ y k , y k 1 ] varying the size of the map and
T

tested with 400 newly generated data samples. Identification of the missile system with
1818 multiple models, which was chosen as the optimal dimension of network as shown
in Figure 5-26(a), is depicted in Figure 5-26(b) where we observe that the models present
a very good approximation of the plant.
Table 5-9. Comparison of modeling performance for the missile system.
Methodology

NRMSE

Multiple Models (1818)

6.1e-3

TDNN (4:35:1)

6.3e-3

88
The modeling performance using the designed multiple models were compared with that
by a TDNN, which was trained by one hidden layer with 35 PEs, in Table 5-9. The
proposed method showed marginally better performance than the nonlinear model did.
Based on 324 linear models and a nonlinear TDNN model with 35 PEs in the
hidden layer, we designed the multiple model-based controllers and the TDNN model
based controller. The optimal parameters selected for each controller were 30 hidden PEs
for TDNN controller, the desired pole locations of 0.3 and 0.7 for the closed loop
response by the PID controller, and c1 = 1 , c 2 = 1.85 , qT = 0.51 for QSMC.

Figure 5-27.

(a)

(b)

(c)

(d)

Tracking various set-point reference signal by (a) TDNNC (b) MIC (c)
MPIDC and (d) MQSMC in the absence of measurement noise.

89
The performance of the proposed control scheme in the absence of noise was
evaluated by comparing it with that of the TDNN controller for the tracking of AOA in
Figure 5-27, in which the MPIDC showed the smallest steady-state error although its
response is sluggish and the MQSMC produced the shortest settling time. Both of them
did not have any overshoot. On the other hand, the TDNNC and the MIC, which have
similar responses, generated unnecessary overshoot; consequently, it took longer than
other controllers to settle down. As a result, the proposed MQSMC is capable of
achieving precise tracking objective with a sufficiently fast response characterized by c .

Figure 5-28.

(a)

(b)

(c)

(d)

Tracking various set-point reference signal by (a) TDNNC (b) MIC (c)
MPIDC and (d) MQSMC under the presence of measurement noise whose
standard deviation is 4.6e-2.

90
In practical implementations of control structures for trajectory control, one
difficulty in achieving accurate trajectory tracking is the existence of observation noise.
To study the effects of this situation, which is very likely to be encountered in practice,
the information used by the controller is corrupted by a random noise having zero mean
and variance equal to 2e-3 and 5.3e-2 for the worst case. The simulation results in Figure
5-28 show that, with appropriate choices of the sliding surface and parameters, robust
tracking can be obtained by the MQSMC even in the presence of sensor noise.
Table 5-10.

Figure 5-29.

Comparison of controller performance for the missile system in the

presence of noise having the standard deviation of 4.6e-2.
Controller

Standard Deviation
of error

TDNNC

10.44e-2

MIC

18.35e-2

MPIDC

10.99e-2

MQSMC

4.40e-2

Trajectory tracking by TDNNC (left) and MQSMC (right) in the presence

of noise whose standard deviation is 2.3e-1.

Additionally, we compared the controller performance for the task of set-point

tracking of AOA starting from 0 degree to 4 degree in the presence of noise having the

91
standard deviation of 4.6e-2 and listed in Table 5-10, in which we observe that the
MQSMC is the only controller which produced less standard deviation of error than the
standard deviation of noise. Figure 5-29 compares the trajectory tracking behavior of the
MQSMC with the TDNNC, in which the MQSMC demonstrates the superior
performance for disturbance rejection compared to its counterpart.

Figure 5-30.

Set-point tracking behavior by the TDNNC (left) and the MQSMC (right)
under parameter variations.

Finally, we regulated AOA to stay 0 degree varying Mach number from 1 to 3 in

order to test the controller performance under parameter variations. From Figure 5-30 it is
clear that the TDNNC suffers from chattering, especially, at the point where the Mach
number is changed, it showed high overshoot. In contrast, the MQSMC demonstrates its
excellent robustness despite parameter variations without any vibration due to its ability
of staying in the switching surface and prompt switching of controllers. Consequently,
the proposed MQSMC provides a robust control system in the presence of either an
external disturbance or parameter variations.

92
5.3.2 LoFLYTE UAV
The second example of the flight vehicle considered for this study is the
LoFLYTE UAV designed by Accurate Automation Corporation (AAC), shown in
Figure 5-31.

Figure 5-31. General description of aircraft (left) and LoFLYTE testbed UAV (right).
The LoFLYTE program is an active flight test program at the Air Force Flight Test
Center at Edwards Air Force Base, with the objective of developing the technologies
necessary to design, fabricate, and flight test a Mach 5 waverider aircraft [18,19]. In
addition, the LoFLYTE UAV has been used to understand the low speed characteristics
of a hypersonic shape and to demonstrate several innovative flight control technologies.
In classical notation, longitudinal motion consists of pitching ( , q ) motion, while
the lateral motion consists of rolling ( , p ) and yawing ( , r ) movement. The elevator
( e ) and the throttle ( th ) control the longitudinal motion, while the aileron ( a ) and
rudder ( r ) primarily affect lateral motion. The general dynamics of the system are
described by

93
u& = ( wq vr ) g sin + F x / M a
v& = (ur wp ) + g cos sin + F y / M a
w& = ( vp uq ) + g cos cos + F z / M a
p& = (( I yy I zz ) qr + I xz ( r& + pq ) + L ) / I xx

(
r& = (( I

q& = ( I zz I xx ) rp + I xz ( r 2 p 2 ) + M / I yy
xx

I yy ) pq + I xz ( p& qr ) + N ) / I zz

(5.8)

& = p + q sin tan + r cos tan

& = q cos r sin
& = q sin sec + r cos sec
where M a is the aircraft mass and Fx , Fy , Fz , L, M , N are actuator-induced forces and
moments. The forces and moments are nonlinear functions of an aircrafts states and
control inputs [98]. This model was programmed in a software simulator by AAC and is
used here as the source of the data. In this study, we wish to estimate and control the
aircrafts lateral motion under the assumption that we can only access roll-rate (p) and
yaw-rate (r), while the goal is to track the desired trajectories ( p d and rd ) during the
course of the flight considering the case of an aircraft moving with a constant throttle.
To model the aircraft lateral dynamics, a total of 2 SOMs are used for quantization

of each embedded output as predictors, y ,k = y k ,L , y k dy +1 , because as the number of

T

inputs to the SOM is increased, the SOM becomes increasingly inaccurate to quantize in
the feature space. Thus the linear coupling between p, q, r , u , v, w is only implicitly
modeled6. In this way, each output (either p or r ) of the aircraft can be described by a
dynamic model that takes into account the control input variables such as a , r :

As we mentioned in Chapter 2, due to difficulties related with dynamic range normalization, local linear
models that take state coupling into account are not as accurate as this approach. Instead, we utilize the
delayed outputs in order to compensate for the disregarded information due to the coupling.

94
y k +1 = f i ( u ,k , y ,k ), i = 1,L , N

(5.9)

where u ,k = { a ,k ,L , a ,k du +1 , r ,k ,L , r ,k du +1 } . By doing this, the complexity can be

T

reduced and it helps to understand the raw data.

Again, we selected an embedding dimension based on the Lipschitz index. Since
the optimal values of embedding dimension for each output dynamic model are different,
we chose the largest number among the embedding dimensions selected for each output
estimation as dy = 2 for each output and du = 6 for 2-D control inputs (aileron and
rudder). The linearized input/output relationship then is
2

i =1

i =1

p k +1 = aip p k i +1 + b1p,i a ,k i +1 + b2p,i r ,k i +1

2

rk +1 = a r
i =1

Figure 5-32.

r
i k i +1

+ b a ,k i +1 + b r ,k i +1
i =1

r
1,i

r
2 ,i

}
(5.10)

Control inputs ( a , r ) used to generate data samples for training the

networks.

The training samples were obtained by exciting the aircraft dynamics (LoFLYTE
Simulator) running at a frequency of 10 Hz using the control inputs shown in Figure 532. Each SOM is trained with the embedded output, y whose dimension is 2, over 5000
samples and the reasonable result for identification of the roll-rate was obtained with a

95
66 grid map ( N = 36 ) as shown in Figure 5-33 where also the modeling performance
with the selected number of models is depicted.

(a)

(b)

Figure 5-33.

Modeling of a roll-rate using multiple models: (a) Identification

performance varying the network dimension (b) Identification of a rollrate ( p ) dynamics by 66 multiple models.

Table 5-11.

Comparison of modeling performance for the lateral motion ( p and r ) of

the LoFLYTE UAV.
Methodology
Multiple models (66)
TDNN (16:50:2)

NRMSE

roll-rate ( p )

yaw-rate ( r )

4.7e-3
7.9e-3

3.8e-3
3.9e-3

After training the SOMs, 36 local linear models were constructed and the created
models were tested by new sequence with 1,000 samples. Table 5-11 shows the
identification performance of two dynamics of the system with 36 models and compares
their performances with that from a TDNN model. Training conditions for a TDNN
model, such as the embedding dimension was kept the same in this comparison between
the local modeling and the global modeling. The best result with a TDNN model was

96
obtained from 20 Monte-Carlo simulations with 50 PEs in the hidden layer7. From the
table, we can conclude that the constructed SOM-based network is a good model of the
underlying dynamics because it provides smaller NRMSE for all dynamics than the
TDNN model. Consequently, it turned out that the proposed strategy of finding proper
location of fixed models depending on the prior information available to the designer for
finding aircraft dynamics is superior to those using a single global nonlinear model. In
addition, it should be noted that the proposed modeling scheme makes identification of
the plant very compact and computationally efficient since the aircraft dynamics are
captured in a compact lookup table of linear models.
We now consider the control problem with the SOM-based local models created.
When we design controllers we usually assume that the coupling between lateral and
longitudinal motion is minimal. Here, we performed a simulation to control the roll-rate
( p ) and yaw-rate ( r ) of the aircraft by aileron ( a ) and rudder ( r ), setting elevator to
zero and throttle to constant. Thus, once we have the linear models for the roll-rate and
the yaw-rate, and the desired values, p d ,k +1 and rd ,k +1 , the inverse controller (inversionbased predictive model), a ,k and r ,k , for the aircrafts roll-rate and yaw-rate tracking is
obtained by
p
, p k , p k 1 , a ,k 1 , L , a , k 5 , r , k 1 , L , r ,k 5
a , k
1 d , k +1
= f i r

,
r
,
r
,

,
L
,

,
L
,

d
k
+
k
k

a
k

a
k

r
k

r
k

,
1
1
,
1
,
5
,
1
,
5
r ,k

(5.10)

And the sliding mode controller was designed such that the poles of the error dynamics
are placed at 0.5 and other parameters are set to qT = 0.5 and rT = 0.001 . Also, for

We tried various sizes of TDNNs and found that the one with 12 PEs in the hidden layer performs best
in system identification.

97
performance comparisons, we applied the TDNN controller, which has 100 PEs in the
hidden layer, for the same control problems.
Figure 5-34 and Figure 5-35 compare the set point tracking performance of the
TDNNC, the MIC, and the MQSMC in the absence of sensor noise and in the presence of
noise whose level is 20 dB of SNR, respectively. From the responses it can be seen that
the multiple controller approach is very good except for the first few seconds. However,
it shows poor transient response when the global control, the TDNNC, is utilized.
Another performance test is to enforce the tracking of the roll-rate and yaw-rate to
signals p d ,k +1 and rd ,k +1 which are given in real time during the course of the flight, while
being subjected to unmeasured sensor disturbances. The 2 output measurements are
corrupted by zero-mean random sequences with 20 dB of SNR. The results of a flight test
with the proposed method are shown in Figure 5-36 where we also show the same with
the TDNNC. It can be seen that the roll-rate and the yaw-rate track their command
signals quite well even under the existence of measurement noise by the multiple
controllers. The simulated flight test demonstrates that the proposed controller is capable
of closely approximating the given mission by only looking at the past information. Also,
it proves that the multiple controller framework indeed provides exceptional tracking.

98

(a)

(b)

(c)
Figure 5-34.

Comparison for controlling roll-rate and yaw-rate to track the set point in
the absence of noise by (a) TDNNC (b) MIC (c) MQSMC.

99

(a)

(b)

(c)
Figure 5-35.

Comparison for controlling roll-rate and yaw-rate to track the set point in
the presence of noise by (a) TDNNC (b) MIC (c) MQSMC.

100

(a)

(b)

(c)
Figure 5-36.

Performance of controlling roll-rate and yaw-rate to track an arbitrary

trajectory with measurement noise (SNR = 20) by (a) TDNNC (b) MIC (c)
MQSMC.

CHAPTER 6
CONCLUSIONS AND FUTURE WORK
In this dissertation, a novel approach to the modeling and control of complex
nonlinear plants was developed. The modeling technique is local and empirical in the
sense that only input-output data from the experiments were used to identify the model,
and the structure of the model is controller orientated. It was shown how the design of
multiple controllers can be based on the model structure presented.
6.1 Summary
The problem of nonlinear system identification and control system design was
addressed under the divide-and-conquer principle. Classical control approaches involving
linearization and gain scheduling in which the parameters of the controller are scheduled
according to changes in system dynamics are seldom applicable to complex nonlinear
systems since there are no general rules for designing gain scheduling regulators and a
general approach. However, an obvious way to analyze complex systems is to adopt the
divide-and-conquer type strategy. This principle motivated the use of multiple local
models for system identification in order to simplify the modeling task. The operating
region is decomposed into smaller sub-regions, which are then described by local models
of simple, possibly linear structure.
Especially in the case of unknown dynamics, where only input-output data from the
plant are available, the proposed method is able to approximate the nonlinear dynamics of
the plant using a piece-wise linear dynamical model that is optimized solely from the
available data. In particular, when local linear models are used as described, it also
101

102
became possible to design a piece-wise linear controller for the plant, whose design is
based on the identified model.
The concept of self-organization was taken in embedded output space extended
with multiple models by the SOM for system identification. The significance of the
proposed scheme is that the operating region is selected by the embedded output and that
local models are built by the embedded output as well as the embedded control input data
samples which are spaced in the local area, which marks the fundamental difference
between this work and others. The effectiveness of the proposed approach was shown
through experiments for modeling complex nonlinear plants such as chaotic systems,
nonlinear discrete time systems and flight vehicles. Its comparison with neural networksbased alternatives (e.g., TDNN) showed clear advantages of local modeling in terms of
performance.
An added advantage of the proposed local linear modeling approach is it greatly
simplifies the design of control systems for nonlinear plants. In general, this is a daunting
task and typically practical solutions involve linearization of the dynamics and then
employing well-established controller design techniques from linear control systems
theory. While designing globally stable nonlinear controllers with satisfactory
performance at every point in the state space of the closed loop control system is
extremely difficult, and perhaps impossible to achieve especially in the case of unknown
plant dynamics, by using the local linear modeling technique presented, coupled with
strong controller design techniques from linear control theory and recent theoretical
results on switching control systems, it becomes possible to achieve this goal through the
use of this much simpler approach of local modeling.

103
The multiple model based linear controllers, such as inverse controller and PID
controller, were introduced for a general class of nonlinear unknown discrete-time
systems via SOM, which divides the state space into a set of operating regions. In
addition, the multiple quasi-sliding mode control strategy was proposed. Contrary to what
is known in the field of sliding mode controller design, the plant dynamics under control
are assumed to be unknown. This is a challenge in the conventional design framework
with the ambiguities introduced by the noise on the measured quantities.
The problems that arise due to the uncertainties of the plant model and
measurement noise are alleviated by incorporating the robustness provided by the sliding
mode technique into the multiple modeling approach. The simulation results
demonstrated that the algorithm proposed is able to compensate deficiencies caused by
the imperfect observations of the state variables and complex plant dynamics, driving the
tracking error vector to the sliding manifold and keeping it on the manifold. In addition,
the proposed method shows better robustness against noise, faster transient response, and
better steady-state accuracy of the controlled system by switching local controllers
astutely through the SOM than other neural network-based alternatives.
6.2 Future Work
The research efforts described in this dissertation were aimed at the development of
a local modeling and control paradigm, and it has been justified by experiments with
various nonlinear plants. During the process of this research work, a number of issues
have arisen to achieve a better modeling and control performance.

The model-controller pair has been selected by looking only at the history of the
plant output in order to avoid some difficulties such as normalization of the inputoutput joint space and large dimensionality of the space involved. The system
dynamics, however, may be captured more effectively in the joint space. Thus,

104
efficient partitioning the operating regions in high dimensional joint space would
be helpful to obtain a more accurate model.

There is no theoretical framework based on which appropriate values for the

parameters in SOM training can be chosen, e.g., initial value for the learning rate
and width of the neighborhood functions, and subsequent rate of decrease and
shrinkage, respectively. Hence, how to more effectively quantize the state space by
SOM remains for further study.

An inverse framework has been employed to design multiple controllers using

identified models. In spite of its simplicity and fast response, some assumptions
should be accompanied: the model is very precise and noise or external
disturbances are negligible. If there is disturbance, then the plant output tracks a
signal which is equal to the desired output plus the disturbance. For this reason,
adding an adaptive scheme or a disturbance canceller, which feeds back an estimate
of the disturbance to the controller of the multiple inverse control strategy, is
required for robust control performance.

To make use of inverse control, the plant must be stable or it must be stabilized.
Real-time stabilization of an unknown, unstable and possibly nonlinear plant is a
difficult problem, and should be explored.

It has been shown that a multiple sliding mode control scheme guarantees BIBO
stability of the overall system. We need to further investigate the stability
conditions for switched control systems.

In sliding mode controller design using multiple models, the sliding hyperplane has
been determined such that the roots of the closed loop response are inside the unit
circle, which guarantees an asymptotic convergence to the desired output. The
closed loop system, however, may have very fast transient response, and possibly
too large an unexpected overshoot depending on the measurement noise level.
Therefore, it would be valuable if an advanced structure for determining sliding
surface considering disturbances is invented.
6.3 Concluding Remarks
We have presented multiple model based control strategy for tracking control of

general unknown nonlinear systems through approximating the state space using a set of
local linear models. The simulation results showed that the derived control using multiple
models turns out to be highly promising and easy to implement in control engineering
practice. These are encouraging results that motivate the use of this modeling and control
technique for various other control applications. The capabilities of the local linear

105
modeling approach is not limited to system identification and control applications. There
are a wide range of nonlinear signal processing problems, such as magnetic resonance
imaging, speech processing, and computer vision, where the concept of local modeling
can be employed to obtain simple but successful solutions to difficult nonlinear problems.

APPENDIX
QSMC FOR MIMO SYSTEM
Consider one of the local l multi-input multi-output models f i of the plant f ,
y1,k +1 = a1,1 y1,k + L + a1,m y1,k m +1 + b1,1,1u1,k + L + b1,1,n u1,k n +1
y l ,k +1

+ L + b1,l ,1u l ,k + L + b1,l ,n u l ,k n +1

M
= al ,1 y l ,k + L + al ,m y l ,k m +1 + bl ,1,1u1,k + L + bl ,1,n u1,k n +1

(A.1)

+ L + bl ,l ,1u l ,k + L + bl ,l ,n u l ,k n +1

where available measurement u, y l and the embedding dimensions of output and

input are m and n , respectively. The state-space model of (A.1) can be written as
r
r
r
r
r
x k +1 = x x + 1u k + 2 u k 1 + L + n u k n +1
and and 1 , L , n have the following forms:
0
0

a1,m
=

O 0
0 1
L a1,1

a1,m 1

0
0

1
O

al ,m

al ,m 1

0
O 0

0 1

L al ,1

0
0
0

b ,L , b
b , L , b
b , L , b
1,l ,1
1,l , 2
1,l , n
1,1,n
1,1, 2
1,1,1

L n =
2 =
M
M
M
1 =

0
0
0

bl ,1,n ,L , bl ,l ,n
bl ,1, 2 , L , bl ,l , 2
bl ,1,1 , L , bl ,l ,1

106

(A.2)

107
The equivalent control produces the following equivalent system
r
r
x k +1 = [ I 1 (C T 1 ) 1 C T ]x k
and the problem is to find

CT

(A.3)

such that the eigenvalues of the matrix

[ I 1 (C T 1 ) 1 C T ] are all inside the unit circle.

For simplicity, we set l = 2 . If the switching surface parameters are defined as

c1,1 L c1,m

CT =
O

0
cl ,1 L cl ,m

(A.4)

then
c1,m b1,1,1
C T 1 =
cl ,m bl ,1,1

c1,m b1,l ,1
cl ,m bl ,l ,1

To simplify the problem we choose

c1,m b1,1,1cl ,m bl ,l ,1 cl ,m bl ,1,1c1,m b1,l ,1 = 1, c1,m = 1, cl ,m = 1 / c1,m (b1,1,1bl ,l ,1 bl ,1,1b1,l ,1 ) .

r
r
This produces x k +1 = [ I 1 (C T 1 ) 1 C T ]x k subject to s k = 0 , in which case

c1,1 L c1,m
r
C T xk =
0

c1,m
0

c 2,1

x1,k
M

0
x m,k

=0
L c 2,m x m +1,k
M

x 2 m ,k

0 x m ,k
c1,1 L c1,m 1
=

c 2,m x 2 m,k
0

c 2,1

x1,k
M

0
x m 1,k

L c 2,m 1 x m +1,k
M

x 2 m 1,k

(A.5)

108
Solving for x m ,k and x 2 m ,k yields

c1,1

x m,k c1,m
x =
2 m ,k

c1,m 1
c1,m

c 2,1
c 2,m

x1,k

M
0
x

m 1,k
c 2,m 1 x m +1,k
L
c 2,m M

x 2 m 1,k

(A.6)

The reduced order equivalent linear system is

0

x1,k +1 0
M c

1,1
x m 1,k +1 c1,m

=
x m +1,k +1
M

x 2 m 1,k +1

O O
0
L

1
c1,m 1
c1,m
0

0
c 2,1
c 2,m

0
x1,k
M

x m 1,k
r
= x k (A.7)
x
1
0 m +1,k
M
O O

0
1 x 2 m 1,k
c 2,m 1
L
c 2,m

The characteristic polynomial of the equivalent system is ( ) = det(I ) . Given

the desired characteristic polynomial as ( ) = 2 m 2 + 12 m 1 + L + 2 m 2 , we can
compute C T by comparing ( ) with ( ) .

LIST OF REFERENCES
1.

Ahmed, M.S., Design of a Fixed Neural Controller for Nonlinear Plants,

International Journal of Control, 69(1):65-83, 1998.

2.

Bartoszewicz, A., Discrete-Time Quasi-Sliding-Mode Control Strategies, IEEE

Trans. Industrial Electronics, 45(4):633-637, 1998.

3.

Bottou, L. and V. Vapnik, Local Learning Algorithms, Neural Computation,

4(6):888-900, 1992.

4.

Brown, R.E., G.N. Maliotis, and J.A. Gibby, PID Self-Tuning Controller for
Aluminum Rolling Mill, IEEE Trans. Industry Applications, 29(3):578-583, 1993.

5.

Camps, J., F.L. Lewis, and R. Selmic, Backlash Compensation with Filtered
Prediction in Discrete Time Nonlinear Systems by Dynamic Inversion Using
Neural Networks, Proceedings of the IEEE Conf. on Decision and Control, 4:
3354-3540, 2000.

6.

Cao, L., Practical Method for Determining the Minimum Embedding Dimension
of a Scalar Time Series, Physica D, 110(1):43-50, 1997.

7.

Casdagli, M., Nonlinear Prediction of Chaotic Time Series, Physical Review D,

35(3):35-356, 1989.

8.

Chakraborty, G. and B. Chakraborty, A Novel Normalization Technique for

Unsupervised Learning in ANN, IEEE Trans. Neural Networks, 11(1):253-257,
2000.

9.

Chan, C.Y., Robust Discrete Quasi-Sliding Mode Tracking Controller,

Automatica, 31(10):1509-1511, 1995.

10.

Chen, G., G. Chen, and R.J.P. Figueiredo, Feedback Control of Unknown Chaotic
Dynamical Systems Based on Time-Series Data, IEEE Trans. Circuits and
Systems-I, 46(5):640-644, 1999.

11.

Chen, L. and K.S. Narendra, Intelligent Control Using Neural Networks and
Multiple Models, Proceedings of the American Control Conference, 2:1357-1362,
2002.

109

110
12.

Chen, X., T. Fukuda, and K.D. Young, Adaptive Quasi-Sliding-Mode Tracking

Control for Discrete Uncertain Input-Output Systems, IEEE Trans. Industrial
Electronics, 48(1):216-224, 2001.

13.

Cho, J., J. Lan, G. Thampi, J.C. Principe and M.A. Motter, Identification of
Aircraft Dynamics Using a SOM and Local Linear Models, Proceedings of the
IEEE Int. Midwest Symposium on Circuits and Systems, 2:148-151, 2002.

14.

Cho, J., J. Lan, J.C. Principe, and M.A. Motter, Modeling and Control of
Unknown Chaotic Systems via Multiple Models, To be appeared in Proceedings
of the Int. Workshop on Machine Learning for Signal Processing, 2004.

15.

Cho, J., J.C. Principe and M.A. Motter, A Local Linear Modeling Paradigm with a
Modified Counterpropagation Network, Proceedings of the Int. Joint Conference
on Neural Networks, 1:34-38, 2003.

16.

Cho, J., J.C. Principe, and M.A. Motter, Local Hammerstein Modeling Based on
Self-Organizing Map, Proceedings of the Int. Workshop on Neural Networks for
Signal Processing, 1:20-24, 2003.

17.

Cho, J., J.C. Principe, D. Erdogmus, and M.A. Motter, Modeling and Inverse
Controller Design for an Unmanned Aerial Vehicle Based on the Self-Organizing
Map, submitted to IEEE Trans. Neural Networks, 2003.

18.

Cox, C., J. Neidhoefer, R. Saeks, and G. Lendaris, Neural Adaptive Control of

LoFLYTE, Proceedings of the American Control Conference, 4:2913-2917, 2001.

19.

Cox, C., R. Saeks, C. Lewis, W. Sefic, A. Carlton, M. Kocher, and R. Pap, Neural
Control of the LoFLYTE Aircraft, Proceedings of the Int. Joint Conf. on Neural
Networks, 2:1047-1050, 1998.

20.

Diao, Y. and K.M. Passino, Intelligent Fault Tolerant Control Using Adaptive
Schemes and Multiple Models, Proceedings of the American Control Conference,
4:2854-2859, 2001.

21.

Efe, M.O. and O. Kaynak, Identification and Control of a Nonlinear Bioreactor

Plant Using Classical and Dynamical Neural Networks, Proceedings of the IEEE
Int. Symposium on Industrial Electronics, 3:1211-1215, 1997.

22.

Eskinat, E., S. Johnson, and W.L. Luyben, Use of Hammerstein Models in

Identification of Nonlinear Systems, AIChE Journal, 37:255-268, 1991.

23.

Farmer, J.D and J.J. Sidorowich, Predicting Chaotic Time Series, Physical
Review Letters, 59(8):845-848, 1987.

24.

Foss, B.A., I.A. Johansen, and A.V. Sorensen, Nonlinear Predictive Control Using
Local Models Applied to a Batch Fermentation Process, Control Engineering
Practice, 3:389-396, 1995.

111
25.

Furuta, K., Sliding Mode Control of a Discrete Time Systems, Systems &
Control Letters, 14:145-152, 1990.

26.

Furuta, K., VSS Type Self-Tuning Control, IEEE Trans. Industrial Electronics,
40(1):37-43, 1993.

27.

Gao, W. and J.C. Hung, Variable Structure Control of Nonlinear Systems: a New
Approach, IEEE Trans. Industrial Electronics, 40(1):45-55, 1993.

28.

Gao, W., Y. Wang, and A. Homaifa, Discrete-Time Variable Structure Control

Systems, IEEE Trans. Industrial Electronics, 42(2):117-122, 1995.

29.

Gawthrop, P.J. and E. Ronco, Local Model Networks and Self-Tuning Predictive
Control, Technical report: CSC-96001, University of Glasgow, 1996.

30.

Ge, M., M. Chin, and Q. Wang, An Extended Self-Organizing Map for Nonlinear
System Identification, Proceedings of the IEEE Conference on Decision and
Control, 1:1065-1070, 1999.

31.

Ge, S.S. and C. Wang, Adaptive Control of Uncertain Chuas Circuit, IEEE
Trans. Circuits and Systems I, 47(9):1397-1402, 2000.

32.

Ge, S.S., C. Wang, and T.H. Lee, Adaptive Backstepping Control of a Class of
Chaotic Systems, Int. Journal of Bifurcation and Chaos, 10(5):1149-1156, 2000.

33.

Gomez, J.C. and E. Baeyens, Identification of Multivariable Hammerstein

Systems Using Rational Orthonormal Bases, Proceedings of the IEEE Conference
on Decision and Control, 1:2849-2854, 2000.

34.

Gonzalez, J., R. Femat, J. Alvarez-Ramirez, R. Aguilar, and M. Barron, A

Discrete Approach to the Control and Synchronization of a Class of Chaotic
Oscillators, IEEE Trans. Circuits and Systems-I, 46(9):1139-1144, 1999.

35.

He, X and H. Asada, A New Method for Identifying Orders of Input-Output

Models for Nonlinear Dynamic Systems, Proceedings of the American Control
Conference, pp.2520-2523, San Francisco, 1993.

36.

Hui, S. and S.H. Zak, On Discrete-Time Variable Structure Sliding Mode

Control, System & Control Letters, 38:283-288, 1999.

37.

Ishigame, A., T. Furukawa, S. Kawamoto, and T. Taniguchi, Sliding Mode

Controller Design Based on Fuzzy Inference for Nonlinear Systems, IEEE Trans.
Industrial Electronics, 40(1):64-70, 1993.

38.

Jacobs, R.A., M.I. Jordan, S.J. Nowlan, and G.E. Hinton, Adaptive Mixtures of
Local Experts, Neural Computation, 3(1):79-87, 1991.

112
39.

Johansen, T.A. and B.A. Foss, Constructing NARMAX Models Using ARMAX
Models, International Journal of Control, 58(5):1125-1153, 1993.

40.

Kaiser, M., Time-Delay Neural Networks for Control, Proceedings of the IFAC
Int. Symposium on Robot Control, pp.112-118, Capri, 1994.

41.

Kantz, H. and T. Schreiber, Nonlinear Time Series Analysis, Cambridge University

Press: Cambridge, UK, 1997.

42.

Kennel, M.B., R. Brown, and H.D.I. Abarbanel, Determining Embedding

Dimension for Phase-Space Reconstruction Using a Geometrical Construction,
Physical Review A, 45:3403-3411, 1992.

43.

Kim, B.S. and A.J. Calise, Nonlinear Flight Control Using Neural Networks,
Journal of Guidance, Control, and Dynamics, 20(1):26-33, 1997.

44.

Kohonen, T., Self-Organizing Maps, Springer-Verlag: New York, NY, 1995.

45.

Lainiotis, D.G., Partitioning: A Unifying Framework for Adaptive Systems,

Proceedings of the IEEE, 64:1182-1198, 1976.

46.

Lawrence, S., A.C. Tsoi, and A.D. Back, Function Approximation with Neural
Networks and Local Methods: Bias, Variance and Smoothness, Proceedings of the
Australian Conference on Neural Networks, pp.16-21, Canberra, 1996.

47.

Lee, C.H. and M.J. Chung, Gain-Scheduled State Feedback Control Design
Technique for Flight Vehicles, IEEE Trans. Aerospace and Electronic Systems,
37(1):173-182, 2001.

48.

Lee, P.M. and J.H. Oh, Improvements on VSS-Type Self-Tuning Control for a
Tracking Controller, IEEE Trans. Industrial Electronics, 45(2):319-325, 1998.

49.

Leontaritis, I.J. and S.A. Billings, Input-Output Parametric Models for Nonlinear
Systems part I: Deterministic Nonlinear Systems, International Journal of Control,
41(2):303-328, 1985.

50.

Lim, S and J.P. How, Modeling and H Control for Switched Linear ParameterVarying Missile Autopilot, IEEE Trans. Control Systems Technology, 11(6):830838, 2003.

51.

Magill, D., Optimal Adaptive Estimation of Sample Stochastic Processes, IEEE

Trans. Automatic Control, 10(4):434-439, 1965.

52.

Martensson, B., Switching Function Adaptive Stabilization with Stability Margin

, Proceedings of the IEEE Conference of Decision and Control, 2:1561-1562,
1989.

113
53.

Martinetz, T.M., S.G. Berkovich, and K.J. Schulten, Neural-Gas Network for
Vector Quantization and its Application to Time-Series Prediction, IEEE Trans.
Neural Networks, 4(4):558-569, 1993.

54.

Maybeck, P.S. and D.L. Pogoda, Multiple Model Adaptive Controller for the
STOL F-15 with Sensor/Actuator Failures, Proceedings of the IEEE Conference
on Decision and Control, 2:1566-1572, 1989.

55.

Maybeck, P.S. and R.D. Stevens, Reconfigurable Flight Control via Multiple
Model Adaptive Control Methods, Proceedings of the IEEE Conference on
Decision and Control, 6:3351-3356, 1990.

56.

Middleton, R.H., G.C. Goodwin, D.J. Hill, and D.Q. Mayne, Design Issues in
Adaptive Control, IEEE Trans. Automatic Control, 33(1):50-58, 1988.

57.

Milosavljevic, C., General Conditions for the Existence of a Quasi-Sliding Mode

on the Switching Hyperplane in Discrete Variable Structure Systems, Automation
and Remote Control, 46(3):307-314, 1985.

58.

Morse, A.S., Toward a Unified Theory of Parameter Adaptive ControlTunability, IEEE Trans. Automatic Control, 35(9):1002-1012, 1990.

59.

Morse, A.S., D.Q. Mayne, and G.C. Goodwin, Applications of Hysteresis

Switching in Parameter Adaptive Control, IEEE Trans. Automatic Control,
37(9):1343-1354, 1992.

60.

Moshou, D. and H. Ramon, Extended Self-Organizing Maps with Local Linear

Mappings for Function Approximation and System Identification, Workshop on
Self-Organizing Maps, pp.181-186, Finland, 1997.

61.

Murray-Smith, R. and K.J. Hunt, Local Model Architectures for Nonlinear

Modeling and Control, in Neural Network Engineering in Dynamic Control
Systems (K.J. Hunt, G.R. Irwin, and K. Warwick, eds.), pp.61-82, Springer-Verlag:
New York, NY, 1995.

62.

Murray-Smith, R. and T.A. Johansen, Multiple Model Approaches to Modeling and

Control, Taylor & Francis: London, UK, 1997.

63.

Narendra, K.S. and C. Xiang, Adaptive Control of Discrete-Time Systems Using

Multiple Models, IEEE Trans. Automatic Control, 45(9):1669-1686, 2000.

64.

Narendra, K.S. and J. Balakrishnan, Adaptive Control Using Multiple Models,

IEEE Trans. Automatic Control, 42(2):171-187, 1997.

65.

Narendra, K.S. and K. Parthasarathy, Identification and Control of Dynamical

Systems using Neural Networks, IEEE Trans. Neural Networks, 1(1):4-27, 1990.

114
66.

Narendra, K.S., J. Balakrishnan, and M.K. Ciliz, Adaptation and Learning Using
Multiple Models, Switching, and Tuning, Control Systems Magazine, 15(3):3751, 1995.

67.

Nelles, O., Nonlinear System Identification, Springer-Verlag: New York, NY,

2001.

68.

Packard, A., Gain Scheduling via Linear Fractional Transformations, Systems

and Control Letters, 22(1):79-92, 1994.

69.

Packard, N.H., J.P. Crutchfield, J. D. Farmer, and R. S. Shaw, Geometry from a

Time Series, Physical Review Letters, 45:712-716, 1980.

70.

Palizban, H.A., A.A. Safavi, and J.A. Romagnoli, A Nonlinear Control Design
Approach Based on Multi-Linear Models, Proceedings of the American Control
Conference, 5:3490-3494, 1997.

71.

Patra, A. and H. Unbehauen, Nonlinear Modeling and Identification, Proceedings

of the Int. Conference on Systems, Man and Cybernetics, 3:441-446, 1993.

72.

Pieper, J.K. and K.R. Goheen, Discrete-Time Sliding Mode Control via InputOutput Models, Proceedings of the American Control Conference, pp.964-965,
San Francisco, 1993.

73.

Poznyak, A.S., W. Yu, and E.N. Sanchez, Identification and Control of Unknown
Chaotic Systems via Dynamic Neural Networks, IEEE Trans. Circuits and
Systems I, 46(12):1491-1495, 1999.

74.

Principe, J.C. and L. Wang, Nonlinear Time Series Modeling with SelfOrganizing Feature Maps, Proceedings of Neural Networks for Signal Processing,
pp.11-20, Cambridge, MA, 1995.

75.

Principe, J.C., L. Wang and M.A. Motter, Local Dynamic Modeling with SelfOrganizing Maps and Applications to Nonlinear System Identification and
Control, Proceedings of IEEE, 86(11):2240-2258, 1998.

76.

Rhodes, C. and M. Morari, Determining the Model Order of Nonlinear

Input/Output Systems Directly from the Data, Proceedings of the American
Control Conference, 3:2190-2195, 1995.

77.

Ritter, R., T. Martinetz, and K. Schulten, Topology-Conserving Maps for

Learning Visuomotor Coordination, Neural Networks, 2:159-168, 1989.

78.

Rivals, I. and L. Personnaz, Nonlinear Internal Model Control Using Neural

Networks: Applications to Processes with Delay and Design Issues, IEEE Trans.
Neural Networks, 11(1):80-90, 2000.

115
79.

Rugh, W.J., Analytical Framework for Gain Scheduling, IEEE Control Systems
Magazine, 11(1):79-84, 1991.

80.

Sales, K.R. and S.A. Billings, Self-Tuning Control of Nonlinear ARMAX

Models, International Journal of Control, 51(4):753-769, 1990.

81.

Sarpturk, S.Z., Y. Istefanopulos, and O. Kaynak, On the Stability of DiscreteTime Sliding Mode Control Systems, IEEE Trans. Automatic Control,
32(10):930-932, 1987.

82.

Sbarbaro, D. and R. Murray-Smith, Self-Tuning Control of Non-linear Systems

Using Gaussian Process Prior Models, DCS Technical report: TR-2003-143,
University of Glasgow, 2003.

83.

Shamma, J.S. and M. Athans, Gain Scheduling: Potential Hazards and Possible
Remedies, IEEE Control Systems Magazine, 12(3):101-107, 1992.

84.

Sidorowich, J.J., Modeling of Chaotic Time Series for Prediction, Interpolation,

and Smoothing, Proceedings of the IEEE Int. Conference on Acoustics, Speech,
and Signal Processing, 4:121-124, 1992.

85.

Singer, A.C., G.W. Wornell, and A.V. Oppenheim, Codebook Prediction: A

Nonlinear Signal Modeling Paradigm, Proceedings of the IEEE Int. Conference on
Acoustics, Speech, and Signal Processing, 5:325-328, 1992.

86.

Sira-Ramiraz, S., Non-Linear Discrete Variable Structure Systems in QuasiSliding Mode, International Journal of Control, 54:1171-1187, 1991.

87.

Stark, J., D.S. Broomhead, M.E. Davies, and J. Huke, Takens Embedding
Theorems for Forced and Stochastic Systems, Nonlinear Analysis: Theory,
Methods & Applications, 30(8):5303-5314, 1997.

88.

Takens, F., Detecting Strange Attractors in Turbulence, Dynamical Systems and

Turbulence (Springer Lecture Notes in Mathematics), 898, pp.365-381, 1980.

89.

Thampi, G., J.C. Principe, J. Cho, and M.A. Motter, Adaptive Inverse Control
Using SOM based Multiple Models, Portuguese Conference on Automatic
Control, pp.278-282, Aveiro, 2002.

90.

Thampi, G., J.C. Principe, M.A. Motter, J. Cho, and J. Lan, Multiple Model Based
Flight Control Design, Proceedings of the IEEE Int. Midwest Symposium on
Circuits and Systems, 3:133-136, 2002.

91.

Tim Sauer, Time Series Prediction by Using Delay Coordinate Embedding, Santa
Fe Institute Studies in the Sciences of Complexity, Addison-Wesley: Reading, MA,
1993.

116
92.

Uduehi, D., A. Ordys, and M. Grimble, Predictive PID Controller Design,

Proceedings of the IEEE Int. Conf. on Control Applications, pp.612-617, Mexico
City, 2001.

93.

Utkin, V., Variable Structure Systems with Sliding Modes, IEEE Trans.
Automatic Control, 22(2):212-222, 1977.

94.

Vesanto, J., Using the SOM and Local Models in Time-Series Prediction,
Workshop on Self-Organizing Maps, pp.209-214, Finland, 1997.

95.

Vu, K.M., Optimal Setting for Discrete PID Controllers, Proceedings of the IEE,
139(1):31-40, 1992.

96.

Walker, D.M., N.B. Tufillaro, and P. Gross, Radial-Basis Models for Feedback
Systems with Fading Memory, IEEE Trans. Circuits and Systems, 48(9):11471151, 2001.

97.

Walter, J., H. Ritter and K. Schulten, Nonlinear Prediction with Self-organizing

Maps, Proceedings of the Int. Joint Conference on Neural Networks, 3:589-594,
1990.

98.

Wan, E.A. and A.A. Bogdanov, Model Predictive Neural Control with
Applications to a 6 DoF Helicopter Model, Proceedings of the American Control
Conference, 1:488-493, 2001.

99.

Wang, C. and S.S. Ge, Adaptive Backstepping Control of Uncertain Lorenz

System, Int. Journal of Bifurcation and Chaos, 11(4):1115-1119, 2001.

100.

Wang, C. and S.S. Ge, Synchronization of Uncertain Chaotic Systems via

Adaptive Backstepping, Int. Journal of Bifurcation and Chaos, 11(6):1743-1751,
2001.

101.

Wang, J. and W.J. Rugh, Feedback Linearization Families for Nonlinear

Systems, IEEE Trans. Automatic Control, 32(10):935-940, 1987.

102.

Wood, G.D., P.J. Goddard, and K. Glover, Approximation of Linear ParameterVarying Systems, Proceedings of the IEEE Conference on Decision and Control,
1:11-13, 1996.

103.

Wu, C.W., T. Yang, and L.O. Chua, On Adaptive Synchronization and Control of
Nonlinear Dynamical Systems, Int. Journal of Bifurcation and Chaos, 6(3):455472, 1996.

104.

Wu, F., A. Packard, and G. Balas, LPV Control Design for Pitch-axis Missile
Autopilots, Proceedings of the IEEE Conference on Decision and Control, 1:188191, 1995.

117
105.

Xianfeng, N., M. Verhaegen, A.J. Krijgsman, and H.B. Verbruggen, A New

Method for Identification and Control of Nonlinear Dynamic Systems,
Engineering Applications of Artificial Intelligence, 9(3):231-243, 1996.

106.

Yang, T. and L.O. Chua, Generalized Synchronization of Chaos via Linear

Transformations, Int. Journal of Bifurcation and Chaos, 9(1):215-219, 1999.

107.

Yang, T., C. Yang, and L. Yang, Detailed Study of Adaptive Control of Chaotic
Systems with Unknown Parameters, Dynamics and Control, 8(3):255-267, 1998.

108.

Yu, X., Tracking Inherent Periodic Orbits in Chaotic Dynamic Systems via
Adaptive Variable Time-Delayed Self Control, IEEE Trans. Circuits and Systems
I, 46(11):1408-1411, 1999.

109.

Zeng, Y and S.N. Singh, Adaptive Control of Chaos in Lorenz System,

Dynamics and Control, 7(2):143-154, 1997.

110.

Zheng, Q. and E. Zafiriou, Control-Relevant Identification of Volterra Series

Models, Proceedings of the American Control Conference, 2:2050-2054, 1994.

BIOGRAPHICAL SKETCH
Jeongho Cho was born in Seoul, Korea, on January 5, 1973. He received his B.S.
degree from Soonchunhyang University, Korea, majoring in control and instrumentation
engineering in 1995. And he earned M.S. degrees in electrical engineering, with emphasis
on systems and control, from Dongguk University, Korea, and the University of Florida
in 1997 and 2001, respectively. Since 2001, he has been with the Computational
NeuroEngineering Laboratory at the University of Florida to pursue his Ph.D. degree. His
research interests include time-series prediction and nonlinear system identification, with
applications to navigation and control.

118