Abstract: Based on the genetic algorithms(GAs), an automatic design and
optimizing method called GMNN (Genetic Multilayer Neural
Network) for feedforward multilayer neural networks is put
forward. Based on the analysis of historical structure design
methods of neural networks, the special genetic encoding
representations are discussed, and a fitness function is defined
also. Some special genetic operations are designed according to
the problem. In addition, simulated annealing algorithms, BP
algorithms and niche technology is also used to accelerate the
convergence rate and improve the generalization of the solution.
The elementary results of the experiments indicate that this
method can find the satisfied network comparatively fast and the
generalization performance of the network that we got is good.
In the process of the software development, Java language is
adopted considering the underlying parallel of the Gas and neural
networks, and the realizing methods of parallel evolution neural
networks are discussed also.
Key words: Artificial
neural network BP model Evolutionary computation Genetic
algorithms
*
This project
is partly supported by National Natural Science Foundation of China (69674037
and 79825102)
|