摘要
BP前馈网络,应用的最为广泛的神经网络,目前拥有许多算法。研究了Matlab神经网络工具箱提供的多种BP算法函数各自的优势和不足,以能在各种不同条件面对不同问题时选用更合适更快速的算法。在介绍了这些算法的基本原理的基础上,以一级倒立摆模型为例进行仿真研究。分别选用简单网络和复杂网络,并对学习步长做了改变,对比了各种BP算法在不同情况下的迭代次数和仿真时间,验证了新型BP算法的优势,得出了对简单网络和复杂网络应该如何选用BP算法的结论。
BP feed forward network, the most widely used neural network, has many algorithms at present. Advantages and disadvantages of various BP algorithms provided in Maflab neural network toolbox are studied so that people can choose more suitable and faster algorithms when different conditions and different problems are faced. After introducing the basic principle of these algorithms, study of simulation is carried out by using a single inverted pendulum as example. Choosing simple and complex nets respectively, and changing the learning steps, the iteration steps and simulation time of various BP algorithms in different conditions are compared. Advantages of new BP algorithms are validated. Advice on how to select BP function is given.
出处
《计算机仿真》
CSCD
2006年第5期142-144,共3页
Computer Simulation
