Studies on genetic multilayer feedforward neural networks and the development of GMNN

Based on genetic algorithms (GAs), an automatic optimizing method for feedforward multilayer neural networks is put forward. The method gives a new genetic encoding representation of the structure of feedforward networks, and a fitness function is defined also. Following the fitness function and some special evolving rules, such as repeating cross operator, we search for the satisfied structure from the network topology space. Based on this method, we develop a simulated program called GMNN (genetic multilayer neural networks). Compared with standard techniques for topology optimization, such as optimal brain surgeon (OBS), magnitude based pruning (MbP) and unit-OBS etc, we concluded that our approach is currently better than other optimization techniques. Some evaluations using GMNN are also given, in these evaluations, we compared the performance of the networks constructed by GMNN to full connection networks, and find that GMNN is better than full connection networks