稳健二进前向网络的遗传训练方法

On Genetic Training of the Robust Binary Feedforward Neural Networks *

稳健二进前向网络是对逻辑知识的隐式表示和显式表示的有机统一，是性能优良的逻辑知识库、推理机和解释器，但目前还没有一种网络训练算法能够训练出稳健二进前向网络，针对这种情况，本文首先对稳健二进前向网络的神经元辅阈值范围进行了改进和有效描述，并在此基础上提出了一种对稳健二进前向网络进行有效训练的遗传训练方法。在这种训练方法中，网络参数采用三值编码方案，并运用相应的变异机制和有效的适应度函数，经这种方法训练出的网络具有结构最优、稳健性能最强和最易实现的特点。文章最后给出了实验的结果。

Robust binary feedforward neural networks(RBFNNs) are the unification of implicit and explicit expression for logic knowledge, and unification of logic knowledge database, inference machine, and interpreter with high performance. However, there is no method to train networks to be RBFNNs at present. For this situation, this paper, after dealing with the complement biase of each neuron in the network, resulting in the fact that it is the sum of the contribution, being-1,0 or 1, of each input to the neuron, suggests a genetic training algorthm for the training of the RBFNNs. In the paper the parameters of the network are coded in three values, and the corresponding crossover operator, mutation operator and various type of fitness functions are utilized. The network trained with the algorithm is of optimal structure, the strongest robustibility, and the easiest realization. Finally, an experiment is presented to show the effectiveness of the algorithm.