Research On Particle Swarm Optimization Algorithms Based On Dynamic Strategies

Particle Swarm Optimization (PSO) is a new paradigm based on swarm intelligence. This paradigm is inspired by concepts from social psychology and artificial life. It simulates social interactions among individuals, namely how individuals imitate behaviors of others in the same group that seem more successful. Its rapid convergence, convenient implementation and small computational requirements make it a good candidate for solving optimization problems. As far, it has become a new focus of the research work in intelligent computation area.Some improved versions of PSO have reported huge success when applied on some problems or in some field. However, the range is very limited, in which some successful variant can be applied. If one wants better result achieved, he must firstly configure the original algorithm according to the problem it is to be applied to solve. In order to eliminate this obstacle, an attempt to design PSOs with dynamic strategies is implemented in this thesis. The proposed algorithms automatically adjust their update processes to achieve ideal results, avoiding the manually configuration process. The main contributions are as follows:A design framework for PSO is constructed and optimized. PSO has been successfully improved and applied in a number of ways since its emergence. However, there has as yet been no uniform framework representing exactly what is involved in modern implementations of the technique. In this thesis a design framework for universal use is constructed by summarizing implementing methods in applications and optimized through synthesizing main achievements in modern PSO. This framework can be considered as a new platform for performance testing of improvements to the technique, as well as a uniform representation in wider optimization community. Subsequent work in this thesis is also carried out based on this platform.A PSO with dynamic memory Dynamic Memory PSO (DMPSO) is designed. Since PSO has imitated the emergence of social norms, based on observation of social phenomenon from the perspective of social psychology, it is studied that how memory will influence individual performance. Then individual memory is defined as a contributor to its success, with several alternative measures provided. A DMPSO is implemented by dynamically assigning proper weight to each individual's memory according to the value of chosen measure formula. The numerical experiment results show that DMPSO can effectively adjust the weight of individual memory, hence automatically adapting to the problem to be solved.Two PSOs with dynamic neighborhood are proposed as Dynamic Neighborhood PSO (DNPSO) and Individually Dynamic Neighborhood PSO (IDNPSO). Analysis of neighborhood topologies and their influence on algorithm performance demonstrates that different problems need different neighborhood topologies to achieve feasible results. Two dynamic neighborhood strategies, one is based on the whole swarm, imitating social reform; and the other is designed from the perspective of individual, similar to the situation that one's communication network changes, are developed to improve the algorithm performance. These two strategies are implemented separately in DNPSO and IDNPSO. The numerical experiment results demonstrate that PSOs with dynamic neighborhood can effectively avoid the pre-mature phenomenon in traditional PSO, as well as generate proper neighborhood topology by automatically adjusting neighborhood structure of a particle swarm.