摘要
分析传统遗传算法易早熟收敛的主要原因,提出一类改进的遗传算法以及一种基于改进遗传算法的前馈神经网络设计方法,用以同时完成对网络结构空间和权值空间的搜索。该算法将普通遗传算法的交叉算子和遗传算子进行改进,利用模拟退火算法、BP算法和小生境技术来加快算法的收敛速度,改善解的性能。通过对异或(XOR)、噪声模式识别等前馈神经网络性能的一组测试,与BP算法进行比较,实验结果表明,该算法能够有效抑制遗传算法初期收敛的发生,有效地提高多层前馈神经网络的收敛精度和收敛速度,由此得到的神经网络的泛化能力也较好,能够达到根据训练样本自动优化设计多层前馈式神经网络的目的,并可获得更为简洁的网络结构。
The primary reason that caulked traditional genetic algorithms to be premature convergence was analyzed. In order to avoid being premature, an improved genetic algorithm was presented. Genetic algorithms （GAS） were used to develop into automatic optimizing method called GMNN （genetic multilayer neural network） for feed- forward multilayer neural networks. The method simultaneously searched the satisfied structure and weights of the network. The method ameliorated the cross operator and the mutation operator of traditional genetic algorithms. Simulated annealing algorithms, BP algorithms and niche technology were used to quicken the velocity of convergence and improve the characteristic of .solution in this method. Through a group of tests to the performance of feed - forward multilayer neural networks such as XOR, noise model identification etc, it was compared with BP algorithms The results of experiments indicate that this method can effectively restrain the premature convergence in genetic algorithm and improve the precision and velocity of convergence for feed - forward multilayer neural networks and the network has good generalization performance. In general, GMNN can automatically design and optimize neural networks by using the training sets and produce more compact artificial neural networks.
出处
《石油化工高等学校学报》
EI
CAS
2006年第1期84-88,共5页
Journal of Petrochemical Universities
关键词
人工神经网络
早熟收敛
BP模型
遗传算法
变异算子
Artificial neural network
Premature convergence
BP model
Genetic algorithms
Mutation operator
