The performance of a modified MEBML system in a noisy environment

Mind-evolution based machine learning (MEBML) is a new evolution learning method that was proposed recently by the authors (1998). MEBML substitutes similartaxis and dissimilation for crossover and mutation operators used in GA. MEBML can solve numerical problems as well as non-numerical ones. In this paper, a new form of similartaxis, called fitted similartaxis, is put forward. In fitted similartaxis, individual´s data in the groups are fitted and the positions of new winners of groups are estimated. Using the least square method in the process of fitting, the influence of noise in the target function is eliminated and the speed of similartaxis is improved. In MEBML, the improvements of speed of similartaxis and dissimilation both help to improve the convergence of the algorithm. Experiments show that MEBML with fitted similartaxis can get high accurate solution of global optima without increasing much of the calculating cost