Breeding Swarms: A GA/PSO Hybrid 簡明昌

Author and Source Author: Matthew Settles and Terence Soule Source: GECCO 2005, p How to get: (\\nclab.csie.nctu.edu.tw\Repository\Journals- Proceedings\ISO Images\GECCO ) Google with paper name

Outline Introduction Breeding Swarms Result and discussion

Outline Introduction Breeding Swarms Result and discussion

Hybrid PSO/GA Angeline points out that the PSO performs well in the early iterations, but has problems reaching a near optimal solution. Both Eberhart and Angeline conclude that hybrid models of the standard GA and the PSO could lead further advances.

Hybrid PSO/GA The LifeCycle model: Combining PSO,GA and HillClimbers Hybrid Particle Swarm Optimiser with breeding and Subpopulation

The LifeCycle Model The ability of an individual to actively decide about its kind of life form in response to its success in its current environment. To create a self-adaptive search heuristic.

The LifeCycle Model

The PSO model used here is similar to the traditional PSO Velocity update rule: Position update rule:

The LifeCycle Model The LifeCycle GA model uses binary tournament selection and elitism to ensure the survival of the individual with the best fitness. Crossover : Uniform Mutation:

The LifeCycle Model HillClimbers are individuals that refine their candidate solution independently of other individuals by examining the local neighborhood. Stochastic HillClimbers Replace probability

PSO with breeding, subpopulation The model incorporates one major aspect of the standard GA into the PSO, the reproduction. Breeding : reproduction and recombination of genes used in standard GA. Subpopulations: Used in GA models mainly to prevent premature convergence.

PSO with breeding, subpopulation

The PSO model used here is similar to the traditional PSO Velocity update rule: Position update rule:

PSO with breeding, subpopulation The breeding is done by first determining which of the particles that should breed with probability pb. From the pool of marked particles, select two random particles for breeding. The parent particles are replaced by their offspring.

PSO with breeding, subpopulation Recombination and reproduction

PSO with breeding, subpopulation The motivation for introducing sub- populations is to keep the diversity Particles are divided into number of sub- populations. Each subpopulation has its own unique best known optimum.

Outline Introduction Breeding Swarms Result and discussion

Background Lovbjerg incorporated a breeding operator into the PSO. Robinson used the GA to initialize the PSO population. GA and PSO were tested and compared when evolving the weights for a fixed topology RANN

Particle Swarm Optimizer A PSO individual retains the knowledge of where in the search space it performed best, a memory of a past experience. However, PSO may waste resources on poor individuals.

Particle Swarm Optimizer

Genetic Algorithm In a GA, if an individual is not selected for elitism or crossover, the information contained by that individual is in a GA. GAs have trouble finding an exact solution and are best at reaching a near optimal solution.

Genetic Algorithm Tournament selection is used with size 3 Gaussian mutation with mean 0.0 and variance reduced linearly each generation from 1.0 to 0.1 Blended crossover (BLX-α)

Genetic Algorithm In Blended crossover Two parents are selected Each gene in the off spring is then calculated by randomly choosing a position in the range

Breeding Swarms Combine the standard velocity and position update rules of PSOs with the ideas of selection, crossover and mutation from GAs An additional parameter, the breeding ratioΨ - D etermines the proportion of the population which undergoes breeding in the current generation

Breeding Swarms In each generation, the bottom (N *Ψ) are discarded and removed from the population. The remaining individual’s velocity vectors are updated, acquiring new information from the population.

Breeding Swarms (N*(1- Ψ)) individuals in the next generation is created by updating the position vectors of the original remaining individuals. Another N*Ψindividuals is filled by GA process.

Breeding Swarms The N*Ψindividuals are selected from the individuals whose velocity is updated to undergo VPAC and mutation and the process is repeated.

Breeding Swarm

VPAC Velocity Propelled Averaged Crossover The goal is to create two child particles whose position is between the parent’s, but accelerated away from the parents current direction.

VPAC

Outline Introduction Breeding Swarms Result and discussion

Parameters setting

Test problemes

Initialization

Result In general, both versions of the BS algorithm outperformed the GA in every test case. Likewise, BS was able to improve on its PSO counterpart in every test case. BS with constriction version was able to perform as well as, or better than BS with inertia in all test cases.

Conclusion The BS was able to locate an optimal, or near optimal, solution significantly faster than either GA or PSO. An additional advantage of the BS may be the ability to implement a variable length string to represent potential solutions.