PSO and its variants

Swarm Intelligence Group

Peking University
Outline
 Classical and standard PSO

 PSO on Benchmark Function

 Analysis of PSO_state of art

 Analysis of PSO_our idea

 variants of PSO_state of art

 Our variants of PSO

 Applications of PSO
Classical and standard PSO
 Swarm is better than personal
Classical and standard PSO

Russ Eberhart James Kennedy

Classical
Vid  w  Vid  c1  Rand ()  ( pid  xid )  c2  Rand ()  ( g d  xid ) (1)
xid  xid  Vid (2)
 Vid : Velocity of each particle in each dimension
 i: Particle
 D: Dimension
 W: Inertia Weight
 c1 、 c2 : Constants
 Rand() : Random
 Pid : Best position of each particle
 gd : Best position of swarm
 xid : Current position of each particle in each dimension
 Classical and standard PSO
Vid  w  Vid  c1  Rand ()  ( pid  xid )  c2  Rand ()  ( g d  xid ) (1)
xid  xid  Vid (2)

xid (t  1)
y Vid (t  1)
g d (t )

Vid (t )
pid (t )
xid (t )
x
Flow chart depicting the General PSO Algorithm:
simulation 1

max
y

min
fitness

x
search space
simulation 2

max
y

min
fitness

x
search space
simulation 3

max
y

min
fitness

x
search space
simulation 4

max
y

min
fitness

x
search space
simulation 5

max
y

min
fitness

x
search space
simulation 6

max
y

min
fitness

x
search space
simulation 7

max
y

min
fitness

x
search space
simulation 8

max
y

min
fitness

x
search space
Schwefel's function
n
f ( x)   ( xi )  sin( xi )
i 1

where
 500  xi  500
global minimum
f ( x) = n  418.9829;
xi = 420.9687, i = 1 : n
Evolution － Initialization
Evolution － 5 iteration
Evolution － 10 iteration
Evolution － 15 iteration
Evolution － 20 iteration
Evolution － 25 iteration
Evolution － 100 iteration
Evolution － 500 iteration
 Search result
Iteration Swarm best
0 416.245599
5 515.748796
10 759.404006
15 793.732019
20 834.813763
100 837.911535
5000 837.965771
Global 837.9658
 Standard benchmark functions

1 ） Sphere Function
n
f x    xi , x   5,5
2 n

i 1

2 ） Rosenbrock Function
n 1
 
f x    100 xi 1  xi
i 1

2 2

 xi  1 , x   10,10
2 n

3 ） Rastrigin Function
D
f x    x  10 cos2x   10
2
i i
i 1

4 ） Ackley Function
 1 n 2 
  exp 1  cos2xi , x   32,32n
n
f x   20  e  20 exp  0.2 x
 n i 1   n i 1 
 
Composition Function
Analysis of PSO_state of art
 Stagnation - Convergence
 Clerc 2002
 The particle swarm - explosion, stability, and convergence in a multidimensional
complex space,2002
 Kennedy 2005
 Dynamic-Probabilistic Particle Swarms,2005
 Poli 2007
 Exact Analysis of the Sampling Distribution for the Canonical Particle
Swarm Optimiser and its Convergence during Stagnation,2007
 On the Moments of the Sampling Distribution of Particle Swarm
Optimisers,2007
 Markov Chain Models of Bare-Bones Particle Swarm Optimizers,2007
 standard PSO
 Defining a Standard for Particle Swarm Optimization,2007
 Analysis of PSO_state of art
 standard PSO: constriction factor - convergence
 Update formula

Vid    (Vid  c1  Rand ()  ( pid  xid )  c2  Rand ()  ( g d  xid ) )

xid  xid  Vid

Equivalent
Vid  w  Vid  c1  Rand ()  ( pid  xid )  c2  Rand ()  ( g d  xid )
xid  xid  Vid
Analysis of PSO_state of art
 standard PSO
 50 particles
 Non-uniform initialization
 No evaluation when particle is out of boundary
Analysis of PSO_state of art
 standard PSO
 A local ring topology
Analysis of PSO_state of art
 How does PSO works?
 Stagnation versus objective function
 Classical PSO versus Standard PSO
 Search strategy versus performance
Classical PSO
 Main idea: Particle swarm optimization,1995
 Exploit the current best position
 Pbest
 Gbest
 Explore the unkown space

pbest gbest
Classical PSO
 Implementation
pbest
gbest

wV

pbest gbest

Vid  w  Vid  c1  Rand ()  ( pid  xid )  c2  Rand ()  ( g d  xid ) (1)

xid  xid  Vid (2)
 Analysis of PSO_our idea
 Search strategy of PSO
 Exploitation
 Exploration
Analysis of PSO_our idea
 Hybrid uniform distribution
wV

pbest gbest

Exploitation
wV

Exploration
Analysis of PSO_our idea
Sampling probability density-computable

x(t  1)  x(t )  wV (t )  Z
Analysis of PSO_our idea
Analysis of PSO_our idea
Analysis of PSO_our idea
Sampling probability
wV
Analysis of PSO_our idea
 No inertia part(wV)
Analysis of PSO_our idea
 Inertia part(wV)
Analysis of PSO_our idea
 No inertia part(wV)
Analysis of PSO_our idea
 Inertia part(wV)
Analysis of PSO_our idea
 Difference among variants of PSO

Exploitation Exploration
Probability

Balance
Analysis of PSO_our idea
 What is the property of the iteration?
Analysis of PSO_our idea
 Whether the search strategy is the same or whether
the PSO is adaptive when
 Same parameter(during the convergent process)
 Different parameter
 Different dimensions
 Different number of particles
 Different topology
 Different objective functions
 In different search phase(when slow or sharp
slope,stagnation,etc)
 What’s the change pattern of the search strategy?
Analysis of PSO_our idea
 What is the better PSO on the search strategy?
 Simpler implement
 Using one parameter as a tuning knob instead of two in
standard PSO
 Prove they are equialent when setting some value of parameter
 Effective on most objective functions
 Adaptive
Analysis of PSO_our idea
 Markov chain
 State transition matrix
 Analysis of PSO_our idea
 Random process
 Gaussian process
 Kernel mapping

Covarance matrix Kernel function Search straegy Effective?

Gauss process Objective

Mapping ability
problem
Analysis of PSO_our idea
 the object of our analysis
 search strategy of PSO
 Different parameter sets
 In different dimensions
 Using different number of particles
 On different objective functions
 Fitness evaluation
 Different topology
 Markov or gauss process and kernel function
 Direction to PSO
 Knob PSO
 Analysis of PSO_our idea

w: x1 1
c: x2 2
dim: x3 3
Num: x4 4
5   f  (  i xi   ) P( Exploitation)
Top: x5 6
FEs: x6 7
Fun: x7
Current results
 Variance with convergence
 func_num=1;
 fes_num=5000;
 run_num=10;
 particles_num=50;
 dims_num=30;
Current results
 Variance with dimensions
 func_num=1;
 fes_num=3000;
 run_num=10;
 particles_num=50;
Current results
 Variance with number of particles
 func_num=1;
 fes_num=3000;
 run_num=10;
 dims_num=30;
Current results
 Variance with topology
Current results
 Variance with inertia weight
 Current results
 1. Shifted Sphere Function
 2. Shifted Schwefel's Problem 1.2

4
x 10
4
x 10
5
8
4
6
3

2 4

1 2

0
0
-1
100 -2
100
50 100
50 50 100
0 50
0 0
-50 -50 0
-50 -50
-100 -100
-100 -100
 PSO on Benchmark Function
 3. Shifted Rotated High Conditioned Elliptic Function
 4. Shifted Schwefel's Problem 1.2 with Noise in Fitness

10
x 10 4
x 10
4
5

3 4

2 3

2
1

1
0
100 0
50 100 100
0 50 0
0 -20
-50 50 -40
-50
-60
-100 -100 -80
0 -100
Current results
 Variance with objective functions
 Unimodal Functions

 Multimodal Functions

 Expanded Multimodal Functions

 Hybrid Composition Functions

Current results
 Variance with objective functions
 func_num=1,2,3,4;
 fes_num=3000;
 run_num=5;
 particles_num=50;
 dims_num=30;
variants of PSO_state of art
 Traditional strategy
 Simulated annealing
 Tabu strategy
 Gradient methods
 Adopted from other fields
 Clonal operation
 Mutation operation
 Heuristical Methods
 Advance and retreat
 Structure topology
 Full connection
 Ring topology
Our variants of PSO
 CPSO
 AR-CPSO
 MPSO
 RBH-PSO
 FPSO
Our variants of PSO
 CPSO

保存每一代的全局最优粒子作为第二步中克隆算子的父粒子

n次迭代过后克隆保存的n个全局最优粒子

将所有克隆出来的粒子利用随机扰动变异

基于浓度机制的多样性保持策略进行选择操作
Our variants of PSO
 MPSO
Our variants of PSO
 AR-CPSO
Our variants of PSO
 FPSO
Applications of PSO
Applications of PSO
Applications of PSO