Niching method using clustering crowding

This study analyzes drift phenomena of deterministic crowding and probabilistic crowding by using equivalence class model and expectation proportion equations. It is proved that the replacement errors of deterministic crowding cause the population converging to a single individual, thus resulting in premature stagnation or losing optional optima. And probabilistic crowding can maintain equilibrium multiple subpopulations as the population size is adequate large. An improved niching method using clustering crowding is proposed. By analyzing topology of fitness landscape using hill valley function and extending the search space for similarity analysis, clustering crowding determines the locality of search space more accurately, thus greatly decreasing replacement errors of crowding. The integration of deterministic and probabilistic replacement increases the capacity of both parallel local hill climbing and maintaining multiple subpopulations. The experimental results optimizing various multimodal functions show that, the performances of clustering crowding, such as the number of effective peaks maintained, average peak ratio and global optimum ratio are uniformly superior to those of the evolutionary algorithms using fitness sharing, simple deterministic crowding and probabilistic crowding.