前向神经网络拓扑结构和权值的进化

Evolution of Multilayer Feedforward Networks

基于进化规则 (BvolutionaryProgramming)的方法 ,提出一种进化前向神经网络的新算法。该算法能同时进化网络的拓扑结构和连接权值 (包括阈值 ) ,产生非常紧凑的网络结构 ,并且由于其全局搜索能力能够避免结构的局部最小。仿真结果证明了该算法的有效性。

Based on evolutionary programming, a novel algorithm for designing the topology and weight distributions of feedforward networks is proposed. This algorithm can evolve network architectures and connect weights(including biases) simultaneously and produce very compact networks. It can escape from structure local minima due to its global search capability. The effectiveness of this algorithm has been confirmed by the simulation results.