Self organized parallel genetic algorithm to automatically realize diversified convergence

In multimodal optimization, maintaining population diversity is one of the most critical issues in genetic algorithm design. A number of niching techniques have been developed and successfully applied to cope with this problem. For multi-population based parallel genetic algorithms, nevertheless, these approaches are obviously inapplicable, since it is very difficult to obtain global information about entire population during parallel evolution procedure. In the present study, a new island model is proposed to overcome this problem. The new method indiscriminately directs local GAs search with considering the topological information of island model. It only uses local information obtained from a few neighbouring subpopulations to achieve a global population diversification. In the new island model, subpopulations are automatically allocated to different regions of searching space so that they could locate multiple optima including both global optima and local optima, simultaneously orders these found optima according to the connection topology of islands, and keeps them until the end of evolution. In addition, through using the proposed method, the performance of PGA is also improved and displays an enhanced global searching capability. Finally, experimental studies, in both unconstrained optimization and combinatorial optimization, are employed to demonstrate the performance of the new island model.