Sei sulla pagina 1di 5

Particle Swarm Optimization

Outline
Voratas Kachitvichyanukul Industrial Engineering & Management Asian Institute of Technology Introduction Properties of particles Design Considerations Strengths and weaknesses Comparison with other methods Applications Q&A

Introduction
developed by Kennedy and Eberhart in 1995. The motivation of PSO algorithm was social behavior such as bird flocking, and fish schooling. PSO is a population-based method, like Genetic algorithm. However, the basic concept is cooperation instead of rivalry.

PSO is very similar to GA, but it does not have genetic (crossover and PSO operators features mutation). A particle moves with the velocity:

its own experience, experience from all particles.

The idea is similar to bird flocks searching for food.


Bird = a particle, Food = the optimal solution pbest = the best solution (fitness) a particle has achieved so far. gbest = the global best solution of all particles

Start

Properties of Particles
1. ability to exchange information with its neighbors 2. ability to memorize a previous position 3. ability to use information to make a decision.

Initialize particles with random position and velocity Evaluate solutions Compare/update pbest

PSO algorithm solution (fitness) a


particle has achieved so far. gbest = the global best solution of all particles.
Yes End

pbest = the best

Compare/update gbest

Meet stopping criterion? No Update velocity and position

Minimization Problem

Minimization Problem Best Particle Other Particle

First Iteration
2

Second Iteration
2 3

Best Particle Other Particle

1. Initializing Position 2. Create Velocity (Vector)

1. Update New Position 2. Create Velocity (Vector)

VELOCITY UPDATING 1. Inertia Term Inertia Term:


3 terms that create velocity: - This new term forces the particle to
2. Cognitive Term move toward in the same direction as before - Audacious tendency, following your own way (just using your old velocity)

3. Social Learning Term Learning Structures Inertia Term

Current Position (X) Personal Best (Pbest)

Velocity Updating
1. Inertia Term

3 terms that create new velocity:


2. Cognitive Term - Conservative tendency 3. Social Learning Term

Cognitive Term: (Personal Best)

-This term forces the particle to go back to the previous best position

c1 * rand ( ) * ( pid xid ) Cognitive Learning

Learning Structures Personal Best (Pbest)

Velocity Updating
1. Inertia Term Social Term:

Current Position (X) Personal Best (Pbest) Global Best (Gbest)

3 terms that create new velocity: -This term forces the particle to
2. Cognitive Term move to the best previous position of its neighbors - Sheep like tendency, be a follower 3. Social Learning Term

c2 * rand ( ) * ( pgd xid ) Social Learning


Learning Structures Global Best (Gbest)

Traditional PSO algorithm Eberhart, R. C. and Kennedy, J. (1995)

vid= vid + c1 * rand()*(pidxid) + c2*rand()*(pgd xid)


Inertia Cognitive Learning Social Learning

Velocity Updating

(1)

vid pid

velocity of dimension d of the ith particle Key Variables best previous position of the ith particle

xid
vid

=
=

xid + vid
[ - Vmax, Vmax ]

(2)

pgd is the best position of the neighbors xid current position of the ith particle c1 & c2 are acceleration constants rand ( ) random function in the range [0, 1] w Inertia weight

Start

Design Considerations
Mapping of particle to solution Number of dimensions Fitness function Number of particles Structure for social learning Values of parameters (c1 & c2 etc.) How to handle infeasible particles Stopping criteria

Initialize particles with random position and velocity Evaluate solutions Compare/update pbest

PSO algorithm

Compare/update gbest

Meet stopping criterion? No Update velocity and position

Yes

End
pbest = the best solution (fitness) a particle has achieved so far. gbest = the global best solution of all particles.

Compare with other heuristics


Evolution algorithm
Simulated annealing (SA) Genetic algorithm (GA) Ant colony optimization (ACO)
SA Prob. search sol. GA Prob. search sol. ACO Prob. construct sol. Population-based
1

GA and PSO similarity


GA
Map sol. to chromosome:
3 2 2 1

JSP
1 3 2

PSO
Map sol. to particle:
2 1

Evolution algorithm
Initiate population Evaluate fitness Generate the next generation

individual solution Population-based

Genetic operator: PSO Prob. search sol. Population-based crossover mutation

Update movement: position velocity

The combination vector created by pbest, gbest, lbest, and nbest pulls each particle to a better GLN-PSO direction than previous published versions
pbest pbest More good information sources, Better performance gbest

Main Components
Intensification is the exploitation of the solutions found in previous searches BALANCE ! Exploitation Diversification is theExploration exploration of the unvisited regions
Quickly find the best solution(s) with in a region Quickly identify region with potentially high quality solution(s)

gbest

Standard PSO

nbest

lbest

GLN-PSO Back

Strength
Simple Easy to maintain diversity via restart or partial restart of particles Easily understood via conceptual comparison with birds flocking

Potential Pitfalls of PSO

Re-initialize some particles

1. Sphere:

Benchmarking Test Functions


3. Rastrigin :
f = Dd=1 (xd2 10cos (2xd) + 10) - Uni-modal - Smooth Surface - High Slope - Multi-modal - Very Rough Surface - Many steep local optimums

f = Dd=1 xd2

2. Rosenbrock:
f = D-1d=1 (100(xd+1 xd2)2 + (xd 1)2)

4. Griewank :
f = 1 + (Dd=1 xd2 /4000) (D d=1 cos (xd /d)) - Multi-modal - Rough Surface - Many slight local optimums

!
- Uni-modal - Smooth Surface - U Shape

Summary
Compounded social learning structures Restart/regeneration of particles Parallel populations with information sharing

Thank you very much

Potrebbero piacerti anche