An improved multi-agent genetic algorithm for numerical optimization

Multi-agent genetic algorithm (MAGA) is a good algorithm for global numerical optimization. It exploited the known characteristics of some benchmark functions to achieve outstanding results. But for some novel composition functions, the performance of the MAGA significantly deteriorates when the relative positions of the variables at the global optimal point are shifted with respect to the search ranges. To this question, an improved multi-agent genetic algorithm for numerical optimization (IMAGA) is proposed. IMAGA make use of the agent evolutionary framework, and constructs heuristic search and a hybrid crossover strategy to complete the competition and cooperation of agents, a convex mutation operator and some local search to achieve the self-learning characteristic. Using the theorem of Markov chain, the improved multi-agent genetic algorithm is proved to be convergent. Experiments are conducted on some benchmark functions and composition functions. The results demonstrate good performance of the IMAGA in solving complicated composition functions compared with some existing algorithms.