PARTICLE SWARM OPTIMIZATION STABILITY ANALYSIS
A Thesis
Submitted to
The School of Engineering of the
UNIVERSITY OF DAYTON
In Partial Fulfillment of the Requirements for
The Degree
Master of Science in Electrical Engineering
by
Ouboti Seydou Eyanaa Djaneye-Boundjou
UNIVERSITY OF DAYTON
Dayton, Ohio
December 2013 PARTICLE SWARM OPTIMIZATION STABILITY ANALYSIS
Name: Djaneye-Boundjou, Ouboti Seydou Eyanaa
APPROVED BY:
Raul´ Ordo´nez,˜ Ph.D. Russell Hardie, Ph.D. Advisor Committee Chairman Committee Member Professor, Electrical and Computer Professor, Electrical and Computer Engineering Engineering
Malcolm Daniels, Ph.D. Committee Member Associate Professor, Electrical and Computer Engineering
John G. Weber, Ph.D. Tony E. Saliba, Ph.D., P.E. Associate Dean Dean, School of Engineering School of Engineering & Wilke Distinguished Professor
ii c Copyright by
Ouboti Seydou Eyanaa Djaneye-Boundjou
All rights reserved
2013 ABSTRACT
PARTICLE SWARM OPTIMIZATION STABILITY ANALYSIS
Name: Djaneye-Boundjou, Ouboti Seydou Eyanaa University of Dayton
Advisor: Dr. Raul´ Ordo´nez˜
Optimizing a multidimensional function – uni-modal or multi-modal – is a problem that reg- ularly comes about in engineering and science. Evolutionary Computation techniques, including
Evolutionary Algorithm and Swarm Intelligence (SI), are biological systems inspired search meth- ods often used to solve optimization problems. In this thesis, the SI technique Particle Swarm
Optimization (PSO) is studied. Convergence and stability of swarm optimizers have been subject of PSO research. Here, using discrete-time adaptive control tools found in literature, an adaptive particle swarm optimizer is developed. An error system is devised and a controller is designed to adaptively drive the error to zero. The controller features a function approximator, used here as a predictor to estimate future signals. Through Lyapunov’s direct method, it is shown that the devised error system is ultimately uniformly bounded and the adaptive optimizer is stable. More- over, through LaSalle-Yoshizawa theorem, it is also shown that the error system goes to zero as time evolves. Experiments are performed on a variety of benchmark functions and results for comparison purposes between the adaptive optimizer and other algorithms found in literature are provided.
iii To my family Gbandi Djaneye-Boundjou, Akoua Maguiloube` Tatcho, Bilaal Djaneye-Boundjou
and Jamila Djaneye-Boundjou.
iv ACKNOWLEDGMENTS
First, I want to thank God for letting me live to see this thesis through. I am thankful to the members of my committee, Dr. Russell Hardie, Dr. Asari Vijayan and Dr. Malcolm Daniels for their time and agreeing to serve as members of my committee. Special thanks to my advisor Dr. Raul´
Ordo´nez˜ for being my advisor, for his exquisite attention to detail and for guiding me throughout
my graduate studies and this research. I am also grateful to Dr.Veysel Gazi, who along with Dr.
Ordo´nez,˜ helped me a great deal while I was working on my thesis.
It is only fitting that I thank Brother Maximin Magnan, SM, and Dr. Amy Anderson without
who I would in all probability not be attending this university in the first place.
Last but not least, I want to thank family and friends for their love, support and encouragements.
v TABLE OF CONTENTS
Page
Abstract ...... iii
Dedication ...... iv
Acknowledgments ...... v
List of Figures ...... viii
List of Tables ...... x
CHAPTERS:
I. INTRODUCTION ...... 1
1.1 Problem Statement and PSO Algorithm ...... 2 1.2 Literature Study on Stability and Convergence ...... 3 1.2.1 PSO Models ...... 3 1.2.2 Velocity Clamping ...... 4 1.2.3 Inertia Factor ...... 5 1.2.4 Theoretical Approach ...... 5
II. MOTIVATION ...... 10
III. STABILITY ANALYSIS ...... 13
IV. TEST FUNCTIONS ...... 25
V. THE ADAPTIVE PSO ...... 28
5.1 Designing the Adaptive PSO ...... 28 5.2 Pseudocode ...... 29 5.3 Understanding the Adaptive PSO ...... 30 5.3.1 Origin-Bias ...... 31 5.3.2 Predictor ...... 33
vi VI. PERFORMANCE EVALUATION ...... 37
VII. CONCLUSION ...... 53
7.1 APSO Characteristics ...... 53 7.2 On Being Biased ...... 54 7.3 Testing the Predictor ...... 54 7.4 APSO Performance ...... 54 7.5 Future Work ...... 55
Bibliography ...... 56
Appendices:
A. MATLAB CODE FOR ADAPTIVE PSO ...... 60
1.1 Main program ...... 60 1.2 Functions used ...... 65 1.2.1 Search space set up ...... 65 1.2.2 Computing the fitness ...... 66 1.2.3 Dead-zone modification ...... 67
B. MATLAB CODE FOR EXPOSING ORIGIN-BIAS ...... 69
2.1 Main program ...... 69 2.2 Functions used ...... 78
C. MATLAB CODE FOR PERFORMANCE COMPARISON TABLE ...... 81
3.1 Main program ...... 81 3.2 Functions Used ...... 86
D. MATLAB CODE FOR FIGURES OF MERIT: CONVERGENCE AND TRANSIENT COMPARISON ...... 87
4.1 Main program ...... 87 4.2 Functions used ...... 100
vii LIST OF FIGURES
Figure Page
4.1 Contour plots of the test functions ...... 26
5.1 Optimizing f1, n = 2 with the search space origin at xo5 ...... 33
5.2 Optimizing f2, n = 2 with the search space origin at xo5 ...... 34
5.3 Optimizing f3, n = 2 with the search space origin at xo5 ...... 34
5.4 Optimizing f4, n = 2 with the search space origin at xo5 ...... 35
5.5 Optimizing f5, n = 2 with the search space origin at xo5 ...... 35
5.6 Optimizing f6, n = 2 with the search space origin at xo5 ...... 36
6.1 Optimizing f1, n = 30 with the search space origin at xo1 ...... 40
6.2 Optimizing f1, n = 30 with the search space origin at xo2 ...... 40
6.3 Optimizing f1, n = 30 with the search space origin at xo3 ...... 41
6.4 Optimizing f1, n = 30 with the search space origin at xo4 ...... 41
6.5 Optimizing f2, n = 30 with the search space origin at xo1 ...... 42
6.6 Optimizing f2, n = 30 with the search space origin at xo2 ...... 42
6.7 Optimizing f2, n = 30 with the search space origin at xo3 ...... 43
6.8 Optimizing f2, n = 30 with the search space origin at xo4 ...... 43
6.9 Optimizing f3, n = 30 with the search space origin at xo1 ...... 44
viii 6.10 Optimizing f3, n = 30 with the search space origin at xo2 ...... 44
6.11 Optimizing f3, n = 30 with the search space origin at xo3 ...... 45
6.12 Optimizing f3, n = 30 with the search space origin at xo4 ...... 45
6.13 Optimizing f4, n = 30 with the search space origin at xo1 ...... 46
6.14 Optimizing f4, n = 30 with the search space origin at xo2 ...... 46
6.15 Optimizing f4, n = 30 with the search space origin at xo3 ...... 47
6.16 Optimizing f4, n = 30 with the search space origin at xo4 ...... 47
6.17 Optimizing f5, n = 30 with the search space origin at xo1 ...... 48
6.18 Optimizing f5, n = 30 with the search space origin at xo2 ...... 48
6.19 Optimizing f5, n = 30 with the search space origin at xo3 ...... 49
6.20 Optimizing f5, n = 30 with the search space origin at xo4 ...... 49
6.21 Optimizing f6, n = 30 with the search space origin at xo1 ...... 50
6.22 Optimizing f6, n = 30 with the search space origin at xo2 ...... 50
6.23 Optimizing f6, n = 30 with the search space origin at xo3 ...... 51
6.24 Optimizing f6, n = 30 with the search space origin at xo4 ...... 51
ix LIST OF TABLES
Table Page
4.1 Test functions ...... 25
∗ 5.1 Impact of moving xa inside Sx ...... 32
6.1 Performance comparison ...... 38
x CHAPTER I
INTRODUCTION
Optimizing a multidimensional function – uni-modal or multi-modal – is a problem that regu- larly comes about in engineering and science. Evolutionary Computation (EC) techniques – includ- ing Evolutionary Algorithm and Swarm Intelligence (SI) – [1, 2, 3, 4, 5, 6] are biological systems inspired search methods often used in solving complex optimization problems. Developed in 1995 by Eberhart and Kennedy [7, 8, 9, 10], Particle Swarm Optimization (PSO) is a proven, powerful, efficient and effective SI technique [11, 12]. PSO is a population-based stochastic optimization method encouraged from social interaction observed in bird flocking or fish schooling. Particles, in a swarm optimizer, are potential solutions to the optimization problem. Their positions and ve- locities are randomly initialized at the start of the search. Each particle then dynamically adjusts its velocity based on previous behaviors and moves about the problem space seeking an optimum solution. At every generation, each particle computes the best solution individually achieved so far, referred to as personal best, and the best solution achieved up to that point by the particles in its neighborhood, referred to as local best, when a local neighborhood topology is employed (not all
particles coexist in the same neighborhood), or global best, when a global neighborhood topology is employed (all particles are neighbors of one another). Particles are attracted towards weighted averages of their personal best and their local best or global best.
1 1.1 Problem Statement and PSO Algorithm
n Let the function f : R → R be a cost function. It is our desire to optimize the cost f using the PSO technique. To do so we can choose from various PSO algorithms studied in literature.
Here, we use the PSO variant with inertia factor, absent from the original algorithm [7, 8]. For i = 1, 2, ... , N, where N is the number of particles in the swarm, the dynamics of particle i, as
described in [13], are given by the equations
i i i i i i i i i v (t + 1) = ω v (t) + ϕ1(t)(p (t) − x (t)) + ϕ2(t)(g (t) − x (t)), (1.1) xi(t + 1) = xi(t) + vi(t + 1),
where all product operations are performed element-wise. The system in (1.1) is a discrete-time
system. The variable t denotes generation or iteration and not continuous time. The position
i i n and velocity vectors of particle i at generation t are respectively x (t), v (t) ∈ R . Moreover,
i i n p (t), g (t) ∈ R are the personal best position achieved by particle i and the local best or global
best position achieved by the neighbors of particle i up to generation t, respectively, such that
pi(t) = arg min f pi(k) , (1.2) k=0,...,t
gi(t) = arg min f pj(k) , (1.3) k=0,...,t j∈Si(k)
with Si(k) being the sets of all the neighbors of particle i at generations k = 0, . . . , t. Further,
i i ω is defined as the inertia factor, ω v (t) is defined as the inertia component, ϕ1 is defined as the
i i i cognitive coefficient, ϕ1(t)(p (t) − x (t)) is defined as the cognitive component, ϕ2 is defined as
i i i i the social coefficient and ϕ2(t)(g (t) − x (t)) is defined as the social component. The vectors ϕ1(t)
i and ϕ2(t) are independent, identically and uniformly distributed n-dimensional random vectors
i i n i i n such that ϕ1(t) ∈]0, ϕ1] and ϕ2(t) ∈]0, ϕ2] . We make sure, for all t, ϕ1(t) 6= 0 and ϕ2(t) 6= 0
for reasons that will become clearer later.
2 1.2 Literature Study on Stability and Convergence
PSO is a valuable optimization technique [14]. Angeline points out that the basic PSO algo- rithm (without the inertia factor), as in [7, 8], helps locates the region nearest to the optima fast but then stales to find it [15]. He thus suggests that adding hybrid components [16] to the basic
PSO algorithm might help it converge faster than other EC techniques. From the first disclosed studies on PSO in 1995 [7, 8], a lot more research has been done on the subject; unsurprisingly, since PSO is easily implementable and can potentially be applied in many ways [17, 18, 19, 20].
Several studies look into how the algorithm’s performance can be improved. Thanks to them, we have an insightful understanding of how parameters such as cognitive and social coefficients, ve- locity clamping (maximum velocity limitation), inertia factor or swarm size affect the algorithm’s performance [14, 20, 21, 22, 23, 24, 25, 26, 27, 28].
1.2.1 PSO Models
Kennedy studies four different models – the cognitive-only model, the social-only model, the full model and the selfless model – of PSO. The full model is as presented in (1.1) but without the inertia factor. The cognitive-only and social-only models omit respectively social component and cognitive component in the velocity update equation. In the selfless model, the local best or global best position of particle i is found by considering only neighboring particles to i while disregarding i itself. Results show that the social-only model generally converges faster than the other models but does so sometimes prematurely and therefore is not as reliable and robust as the full model is [25].
Carlisle and Dozier corroborate Kennedy’s results on dynamic changing environments [23] and propose that picking a larger cognitive coefficient upperbound ϕ1 compared to the social coefficient upperbound ϕ2, such that 0 < ϕ1 + ϕ1 < 4, might even prove to be better [24].
3 1.2.2 Velocity Clamping
At each generation, the velocity of particle i is updated dynamically by adding up its inertia component, its cognitive component and its social component, as shown in the first equation of
(1.1). The second equation in (1.1) shows how the position of particle i is adjusted according to its
updated velocity. The inertia component makes sure a portion of the current velocity information is
contained in the updated velocity. Because of the cognitive component, each particle is able to draw
on and make use of its own experience during the search process. The social component shows that
particles can also use the information received from their neighbors for help throughout the search
process. High cognitive coefficients favor a local search whereas high social coefficients favor a
global search. Summing up all three components can however yield high values of velocity. As a
consequence particles take bigger step sizes when updating their positions, with the risk of them
possibly leaving the search space. Convergence may thus either take longer or not even be possible.
A commonly used approach is to limit velocity within a range [−Vmax, Vmax], with Vmax > 0, such that
i i i v (t + 1), if v (t + 1) ≤ Vmax, v (t + 1) = i i (1.4) sign(v (t + 1)) · Vmax, if v (t + 1) > Vmax, where sign(·) is the sign operator. This approach is called velocity clamping [21]. Picking a large
Vmax can result in particles searching far-removed regions and possibly straying away from regions with good solutions. On the other hand, if Vmax is too small, particles search smaller regions and can easily be trapped in local minima. It should be noted that velocity clamping does not stop par- ticles from escaping the search space. Nevertheless, restricting positions update step sizes prevents, in some way, particles to diverge. A constant Vmax is common procedure. However, adjusting Vmax dynamically might better the optimizer’s performance [29].
4 1.2.3 Inertia Factor
Shi and Eberhart introduce the inertia factor PSO variant [27]. As stated before, the inertia factor controls how much information contained in the velocity at generation t contributes to up- dating the velocity at generation t + 1. Empirical studies suggest that picking the proper inertia factor is necessary to guarantee a convergent optimizer [27, 28, 30]. If ω > 1, velocities grow larger as generations go by and particles diverge from regions with good solutions. If ω < 0 , velocities decrease gradually to zero over time contingent upon the values of the upperbounds ϕ1 and ϕ2. However, choosing the right inertia factor may also depend on the problem at hand [28].
With ω = 0.7298, ϕ1 = ϕ2 = 1.49618, N = 30 and by confining velocities within the same set
as the search space, Shi and Eberhart achieved good optimization results for some objective func-
tions [30]. Trelea also obtained good results for the same objective functions as used in [30] but
without velocity clamping, by setting w = 0.6, ϕ1 = ϕ2 = 1.7 [31]. Thomas Beielstein et al. provide shrewd information about parameters such as inertia factor, swarm size, cognitive coeffi- cient and social coefficient upperbounds by applying design of experiments (DOE) methods to the
PSO algorithm [22]. In general, researchers have successfully experimented with constant inertia factor. However, good results have also been obtained by adaptively changing the inertia factor
[26, 32, 33, 34, 35, 36].
1.2.4 Theoretical Approach
Most aforementioned studies nevertheless only offer empirical evidence. The issue of conver- gence and stability is formally tackled in [13, 31, 37, 38, 39, 40, 41, 42, 43]. A common approach is to analyze either the dynamics or the trajectory of a simplified PSO system so as to derive conditions for convergence and stability [31, 37, 38, 39, 40, 41, 43].
5 Ozcan and Mohan are the first researchers to do so [38, 39]. They analyze a single (N = 1),
scalar (n = 1), deterministic (ϕ1(t) = ϕ1 and ϕ2(t) = ϕ2 such that ϕ1 and ϕ2 are constant for all
t) particle, whose personal best and global best are equal to a constant scalar (p(t) = g(t) = s)
and later extend their findings to a more general swarm optimizer. For i = 1, 2, ... , N and
d = 1, 2, ... , n, where N is the number of particles in the swarm and n is the dimension of
each particle, particles update their positions such that
id id id id id id id x (t) − 2 − ϕ1 − ϕ2 x (t − 1) + x (t − 2) = ϕ1p + ϕ2g . (1.5)
q q id id id id id id2 id 2 id id Let ϕ = ϕ1 + ϕ2 and δ = 2 − ϕ1 − ϕ2 − 4 = (ϕ ) − 4ϕ . By varying ϕ , different cases are studied in [38, 39]. For ϕid = 0, 4 and ϕid > 4, δid is real and particles have a tendency of leaving the search space and never returning. For 0 < ϕid < 4, δid is complex and
solutions to (1.5) are of the form
xid(t) = Λidsin(θidt) + Γidcos(θidt) + κid, (1.6)
id id |δ | id id id where θ = arctan id id and the constants Λ , Γ and κ are given by |2−ϕ1 −ϕ2 |
2vid(0) − ϕid + ϕid xid(0) + ϕidpid + ϕidgid Λid = 1 2 1 2 , (1.7) |δid| Γid = xid(0) − κid, (1.8)
id id id id id ϕ1 p + ϕ2 g κ = id id . (1.9) ϕ1 + ϕ2
Here, as suggested by (1.6) through (1.9), particles surf sine waves whose frequencies and ampli- tudes depend on initial conditions and the choice of the cognitive and social coefficients. What
Ozcan and Mohan analysis has shown is that particles’ trajectories inside the search space are sinu- soidal waves, provided 0 < ϕid < 4. To find an optimum, particles surf from waves to waves by randomly changing their frequencies and amplitudes.
Just like Ozcan and Mohan, Clerc and Kennedy provide a theoretical analysis of the trajectory
6 of a simple particle [37, 43]. They develop a 5-D attractor space that sums up the trajectory of the simple particle and furthers Ozcan and Mohan’s results. Their analysis is based on guaranteeing that particles converge to the stable point
ϕidpid + ϕidgid sid = 1 2 , (1.10) ϕid
id id id with ϕ = ϕ1 + ϕ2 , such that i = 1, 2, ... , N and d = 1, 2, ... , n. As a result of their study,
Clerc and Kennedy derive a constriction factor χid [37], which works just like velocity clamping
by keeping velocities bounded. The new velocity equation is given by
id h id id id id id id id i v (t + 1) = χ v (t) + ϕ1 (t)(p (t) − x (t)) + ϕ2 (t)(g (t) − x (t)) , (1.11) id id 2κ with χ = , (1.12) q 2 2 − ϕid − (ϕid) − 4ϕid
where the constant κid, such that 0 ≤ κid ≤ 1, determines how fast the algorithm converges. As
κid → 0 convergence of particle i along dimension d speeds up and conversely slows down as
κid → 1.
Van Den Bergh and Engelbrech analyze the PSO algorithm variant with inertia factor [40,
41]. They prove that particles converge to a stable point si over time, such that lim xi(t) = t→+∞ i i i i i ϕ1p + ϕ2g s = i i , with all operations performed element-wise, provided for all dimensions d = ϕ1 + ϕ2 n id id o 1, 2, ... , n, max ||λ1 ||, ||λ2 || < 1, where
1 + ωid − ϕid − ϕid + γid λid = 1 2 , (1.13) 1 2 1 + ωid − ϕid − ϕid − γid λid = 1 2 , (1.14) 2 2 q id id id id2 id with γ = 1 + ω − ϕ1 − ϕ2 − 4ω . (1.15)
As a guideline for convergence, Van Den Bergh proves that for all dimensions d = 1, 2, ... , n,
n id id o 1 max ||λ1 ||, ||λ2 || < 1 is achieved in general if ω > 2 (ϕ1 + ϕ2) − 1 [41].
7 A stricter analysis is performed in [42]. Unlike [31, 37, 38, 39, 40, 41, 43], Kardimanathan et al. present a Lyapunov stability analysis of the PSO system considering stochastic parameters.
The absolute stability of the “best” particle is proved and guidelines for picking the inertia factor and maximum gain K for stability are presented. The analysis is performed on a particle swarm optimizer with inertia factor. Because each dimension updates independently of others, the study can be done on a single, scalar particle without loss of generality. Considering a single, scalar particle, and removing the subscripts i and d for clarity, ϕ(t) = ϕ1(t) + ϕ2(t). The maximum gain
is defined as 0 < ϕ(t) < K. The analysis in [40, 41] requires that 0 < ω < 1 and K < 2(ω + 1)
for convergence. However, according to Kardimanathan et al.’s analysis, sufficient conditions for a
stable swarm optimizer considering stochastic parameters are such that |ω| < 1, with ω 6= 0, and
2 1 − 2|ω| + ω2 K < . (1.16) 1 + ω
Gazi refines the findings in [42]. He proves stochastic asymptotic stability of the “best” particle, derives a more relaxed upperbound on the maximum gain 27 1 − 2|ωi| + ωi2 K < , (1.17) 7 (1 + ωi) and shows stochastic global ultimate boundedness of the “non-best” particles [13]. He also proves the convergence of the personal best and global best function value sequences [13].
As Kardikamanathan et al. point out in [42], the main purpose of PSO is optimization, which is only possible if the PSO system remains stable. On one hand, empirical studies focus on devising or finding PSO control parameters for achieving the best possible result but cannot prove stability theoretically. On the other hand, theoretical studies, be it simplified or strict, to a certain extent or fully prove stability. In doing so, guidelines for picking some PSO parameters are derived.
However, in most cases, those guidelines only help achieving stability to the expense of convergence
8 or optimization. Should we just abandon the theoretical approach and computationally find the best
PSO control parameters?
In this thesis, we go the theoretical route. We develop an adaptive PSO based on the work in
[42] and discrete-system adaptive control tools in [44]. The particle dynamics are represented as a nonlinear feedback control system in canonical form. We use Lyapunov’s direct method to prove that our error system is ultimately uniformly bounded and hence the stability of our optimizer. For the chosen benchmark functions, we perform various experiments to further understand how our optimizer work and compare its performance to both previous purely empirical approach results and theoretical approach results.
This thesis is organized as follows: Chapter II describes the motivation behind the study, the stability analysis is done in Chapter III, for the test functions presented in Chapter IV, we investigate our adaptive PSO in Chapter V and compare its performance to other optimizers in Chapter VI, followed by the conclusion of the thesis in Chapter VII.
9 CHAPTER II
MOTIVATION
Considering a global neighborhood topology where all the particles composing the PSO system can exchange information, we can write
g1(t) = g2(t) = ... = gN (t) = g(t). (2.1)
According to (1.1), each of the n dimensions updates independently of others. The cost f is the only link between the dimensions. For each particle, the scalar case can thus be analyzed without loss of generality. Removing the index i for clarity purposes, consider for now a single, scalar particle whose dynamics are given by
v(t + 1) = ωv(t) + ϕ1(t)(p(t) − x(t)) + ϕ2(t)(g(t) − x(t)), (2.2) x(t + 1) = x(t) + v(t + 1).
We base our analysis on Chapter 13 in [44]. Therefore, we proceed by putting (2.2) in the canonical
form described in [44]. Let z1(t) = x(t − 1) and z2(t) = z1(t + 1) = x(t). We can also express
z2(t + 1) as
z2(t + 1) = z1(t + 2) = x(t + 1),
= x(t) + v(t + 1),
= x(t) + ωv(t) + u1(t), (2.3)
10 where, based on [42], we can define the control input signal u1(t) as
u1(t) = ϕ1(t)(p(t) − x(t)) + ϕ2(t)(g(t) − x(t)),
ϕ1(t) + ϕ2(t) = (ϕ1(t)p(t) + ϕ2(t)g(t)) − (ϕ1(t) + ϕ2(t)) x(t), ϕ1(t) + ϕ2(t) = ϕ(t)(s(t) − x(t)), (2.4)
with
ϕ(t) = ϕ1(t) + ϕ2(t) (2.5)
and
ϕ (t)p(t) + ϕ (t)g(t) s(t) = 1 2 . (2.6) ϕ1(t) + ϕ2(t)
Remark that if ϕ(t) = 0 then s(t) in (2.6) and consequently u1(t) in (2.4) become undefined. That
is why we require ϕ1(t) 6= 0 and ϕ2(t) 6= 0 so that ϕ(t) 6= 0 for all t. From (2.2), it can be inferred
that z1(t) = x(t − 1) = x(t) − v(t). Meaning v(t) = x(t) − z1(t). Therefore (2.3) becomes
z2(t + 1) = x(t) + ω(x(t) − z1(t)) + u1(t),
= −ωz1(t) + (1 + ω)z2(t) + u1(t). (2.7)
The single particle system in state-space form is written as z1(t + 1) 0 1 z1(t) 0 = + u1(t), z2(t + 1) −ω 1 + ω z2(t) 1 (2.8) z (t) y(t) = 0 1 1 . z2(t) 0 1 We want A = to be non-singular. Hence, ω 6= 0 since the eigenvalues of A are z −ω 1 + ω z {1, w}. Notice that ω = 0, according to (2.2), means v(t + 1) is independent of v(t), which would
modify the PSO algorithm.
We define an error system
e(t) = s(t) − y(t) = s(t) − z2(t). (2.9)
11 If e(t) can somehow be driven to zero (e(t) → 0) then
ϕ1(t)p(t) + ϕ2(t)g(t) x(t) = z2(t) = s(t) = . (2.10) ϕ1(t) + ϕ2(t)
At any generation t, there always is at least one particle whose personal best is the same as its global best [13]. Such particle is called “best” particle [42]. Remember that the scalar case, given by (2.8) through (2.9), applies to each of the n dimensions and N particles independently. Going back to the n-dimensional, N-particle PSO system, the best particle j (j ∈ [1,N]) is such that pj(t) = gj(t) = g(t). If we can get the error to go to zero, according to (2.10), xj(t) = pj(t) =
j j j j j g (t) = g(t). Moreover, using (1.1), x (t + 1) = x (t) + ω v (t). There exists g ∈ R such that g = lim min gi(t) . Gazi proves that the sequence gi(t) converges [13]. Here, because we t→∞ 1≤i≤N use a global neighborhood topology, we can redefine g as
g = lim g(t). (2.11) t→∞
There is no guarantee that g is the position that globally optimizes the cost f. However, when gj(t) = g, the particle j should remain in the same position at the next iteration, meaning xj(t+1) =
j j j j j j j x (t) = g. In that case, for x (t + 1) = x (t) + ω v (t) to hold, ω v (t) = 0. Recall that for Az to be non-singular, ω 6= 0. Consequently, vj(t) = 0. After all, if xj(t) = g and vj(t) = 0 then the
cost f is optimized and the position at the optimum is xj(t) = gj(t) = g. As mentioned before, we can only hope xj(t) = g is the position of the global optimum. The question now that remains however is, how can the error system e(t) be driven to zero?
12 CHAPTER III
STABILITY ANALYSIS
In this chapter we present a stability analysis based on Lyapunov direct method. It is therefore only right that we define it.
Lyapunov Stability [45]: The equilibrium xe = 0 of a system is stable if there exists a nonneg- ative, scalar Lyapunov function V (t, x), with V (t, 0) = 0, which satisfies
V (t + 1, x) − V (t, x) ≤ 0. (3.1)
LaSalle-Yoshizawa Theorem [44]: Here also, let the xe = 0 be the equilibrium of a system.
Replacing V˙ (t, x) in [44] by V (t + 1, x) − V (t, x), if there exists a nonnegative, decrescent, scalar Lyapunov function V (t, x), with V (t, 0) = 0, such that
V (t + 1, x) − V (t, x) < −W (x) < 0, (3.2)
where W (x) is positive semi-definite then xe = 0 is uniformly bounded and lim W (x) = 0. t→+∞ Now, returning to the single particle case and considering only one of the n dimensions, we
design an adaptive controller to drive e(t) to zero. The original controller u1(t) in (2.4), which can
also be expressed as u1(t) = ϕ(t)e(t), is replaced by
u(t) = u1(t) + U(t), (3.3)
13 with U(t) being the adaptive controller. The new system is now given by
z (t + 1) 0 1 z (t) 0 1 = 1 + u(t), z2(t + 1) −ω 1 + ω z2(t) 1 (3.4) z (t) y(t) = 0 1 1 . z2(t)
From (2.9), we compute e(t + 1) as
e(t + 1) = s(t + 1) − z2(t + 1),
= s(t + 1) − ϕ(t)e(t) + ωz1(t) − (1 + ω)z2(t) − U(t). (3.5)
Let
h(t) = s(t + 1), ϕ (t + 1)p(t + 1) + ϕ (t + 1)g(t + 1) = 1 2 . (3.6) ϕ1(t + 1) + ϕ2(t + 1)
It is uncertain what h(t) is at generation t as it happens at a generation in the future and cannot be computed until we know x(t + 1). However, we need to eliminate h(t) if we want to drive e(t) to
zero. We do so with a universal function approximator H(t, θ) used here as a predictor. Remark that s(t + 1) is random since ϕ1(t + 1) and ϕ2(t + 1) are random. However we can only hope to be able to predict s(t + 1) because particles are attracted towards the best particle(s). That is p(t) → g, g(t) → g and s(t) → g as t → ∞. We assume that there is a linear in the approximator
ˆ p H(t, θ(t)) such that for some ideal parameter θ ∈ R , |H(t, θ) − h(t)| ≤ L for all x ∈ Sx and ˆ v ∈ Sv, with Sx ⊂ R being the search space, Sv ⊂ R being the velocity space and θ(t) being > ∂H(t, θˆ(t)) the parameter estimate of the ideal parameter θ. That means H t, θˆ(t) = θˆ(t). ∂θˆ(t) We also assume that θ is time-invariant. Moreover, we want to make sure the error e stays inside
the error space Se ⊂ R to guarantee x ∈ Sx and v ∈ Sv. The goal here is to design an adaptive
controller which ensures that if e ∈ Be then a proper choice of the controller parameters can be
made so that Be ⊂ Se.
14 The control law is given by
U(t) = Us(t) + H(t, θˆ(t)), (3.7)
with Us(t) being the static controller.
We can derive Us(t) assuming no uncertainty h(t). That is h(t) = 0 and (3.5) becomes
e(t + 1) = −ϕ(t)e(t) + ωz1(t) − (1 + ω)z2(t) − Us(t). (3.8)
Choosing
Us = α(t) − ke(t), such that |k| < 1 and k 6= 0, (3.9) where
α(t) = −ϕ(t)e(t) + ωz1(t) − (1 + ω)z2(t), (3.10)
2 then e(t + 1) = ke(t), |k| < 1. Consider the Lyapunov candidate Vs(t) = (e(t)) . Vs is positive definite, decrescent and radially unbounded. Further,
2 Vs(t + 1) − Vs(t) = (e(t + 1)) − Vs(t),
2 = k Vs(t) − Vs(t),