基于BP网络的权值更新快速收敛算法

Rapid convergence algorithms for weight values updating based on BP network

针对标准BP网络学习算法收敛慢的问题,提出了两种权值更新的快速收敛算法,即基于梯度变化率的快速传递算法和基于梯度方向的弹性传递算法,并在煤矿事故救援游戏式训练系统中进行仿真和比较,让游戏角色根据井下空气成分学习判断危险程度,以便受训人员或仿生机器人采取相应的措施。仿真结果表明,所提算法的收敛时间比标准算法有一定改善。

To solve the slow convergence of standard learning algorithm in BP network, two rapid convergence algorithms were suggested for weight values updating. One is rapid transmission algorithm based on gradient change rate. The other is flexible transmission algorithm based on gradient orientation. The two algorithms were simulated and compared in Game Style Training System for Mine Accident Rescuing. Here the algorithms would help game roles learn to estimate the danger degree according to ingredients of mine air, and then help trainees or biorobots take corresponding actions. The simulating results show that shorter convergence time is taken for the two algorithms than the standard algorithm.