多种群并行进化神经网络的研究及应用

Research and Application on Multigroup Parallel Evolutionary Neural Network

提出一种新的多种群并行遗传算法 (NMPGA) ,并将其作为多层前馈神经网络(MFNNs)的学习算法 ,从而形成一类新的 MFNN模型——多种群并行进化神经网络(MPENNs)。首先 ,对一给定的网络结构 ,随机产生一初始权重的集合 ,这个集合实际上对应着一组具有相同结构但不同权重的神经网络。然后 ,采用 NMPGA对 MFNNs的权重进行进化。最后 ,性能最好的网络被选作目标问题的解。在 NMPGA算法中 ,作者采用浮点数编码来克服传统二进制编码的精度不足问题 ,并设计了专门的杂交算子和变异算子来增强算法性能。实验结果表明 ,MPENNs能成功解决异或问题、三元奇偶问题及成品烟的感官质量评价问题。

A novel multigroup parallel genetic algorithm (NMPGA) is presented as a learning method for multilayer feedforward neural networks (MFNN). Consequently, a new MFNN model multigroup parallel evolutionary neural network (MPENN) is formed. First, for a given network, architecture initial weight sets are generated randomly, which in fact are corresponding to a group of MFNN with the same architecture but different weights. Then the NMPGA is adopted to evolve the weights and biases of all the MFNN. In the end the best network is chosen as a solution to the object problem. In addition, float encoding is introduced to solve the accuracy insufficiency problem of the traditional binary encoding. Furthermore a new crossover operator and a new mutation operator are devised to enhance the algorithm performance. The experimental results show that MPENN can solve the XOR problem, the 3 bit parity problem and the sensory quality assessment of cigarettes successfully.