Particle Swarm Optimization James Kennedy & Russel C. Eberhart
Idea Originator Landing of Bird Flocks Function Optimization Thinking is Social Collisions are allowed
Simple Model Swarm of Particles Position in Solution Space New Position by Random Steps Direction towards current Optimum Multi-Dimensional Functions
First Feedbacks Fast in Uni-Modal Functions Neuronal-Network Training (9h to 3min) Able to compete with GA (overhead) But, Algorithm is based on Broadcasting Multi-modal Function Optimization
Algorithm Updates Storage of individual Best [Kennedy] Move between individual & global Best Constriction Factor [Shi&Eberhart] Tracking Changing Extreme [Carlisle]
Hybrid PSO Breed & Sub-population Combine Adv. of PSO & EA Anal. comparison PSO vs. GA [Angeline] Idea: Increase Diversification
Hybrid Approach - Breeding Steps Select Breeding Population (pb – prob.) Select two random Parents Replace Parents by Offspring Offspring Creation arithmetic crossover for position & velocity
Hybrid Approach – Sub-Popul. Steps Divide into multiple Subpopul. Spread particles over solution space Use Breeding approach Sub-Popul. Selection Breeding over diff. Poul. (psb – prob.)
Hyb. Results Usage of 4 multi-dim. Functions In uni-modal function GA & std. PSO better In multi-modal function hyp. PSO better convergence & solution Subpopulation results in no gains
Conclusion New Research Area First PSO in 1995, First Conf. Last Year Highly accepted Increasing Research & Evol. Comp. Special Can we learn from GA & PSO a improved method with reduced overhead?
Reading Room “Swarm Intelligence” by Kennedy & Eberhart [2001] Bibliography