Population Topologies and Their In uence in Particle Swarm Performance

Particle Swarm Optimization (PSO) is a new paradigm of Swarm Intelligence. Particle swarms are a valuable tool to nd optima in a tness landscape in <n, especially useful when dealing with a high number of dimensions and problems where problem speci c information is non-existent. Its rapid convergence and small computational requirements make it a good candidate for solving optimization problems. This paradigm is inspired by concepts from Social Psychology and Arti cial Life. It simulates social interactions among individuals, namely the emergence of social norms, and how individuals imitate behaviors of others of the same group that seem more successful. The goal of this simulation is to have the individuals cooperate to nd the global optimum of a tness landscape. This thesis focuses on some aspects that have been taken for granted in particle swarm. The rst one concerns the choice of topology: most researchers use the same ones: gbest and lbest. There is a strong relationship between the choice of the social topology and the robustness of the algorithm to premature convergence. This thesis demonstrates the dramatic increase of robustness that a judicious choice of the population topology can have. Another aspect is the choice of the sources of in uence; these are normally the individual's previous success and the best success observed in the group. A new extension is presented, Fully Informed Particle Swarm (FIPS), which combines all the information available in its neighborhood. Results presented here demonstrate that FIPS can outperform the canonical PSO on all problems tested.