一种用于特征层融合识别的回弹全局自适应动量BP算法

Resilient Global-Adaptive Momentum BP Algorithm for Feature Level Fusion Recognition

针对特征层融合识别中全局自适应BP算法存在的收敛速度慢、学习不稳定等问题,基于对动量BP算法的详细分析,提出了一种新的全局自适应BP算法——回弹全局自适应动量BP算法(RGMOBP),该算法具有在误差增大时进行权值回弹并减小学习步长以保证权值的调节功能、使误差减小的特点。仿真结果表明:RGMOBP在学习性能上优于其它已有经典全局算法,是一种简单有效的神经网络特征层融合识别算法。

The global-adaptive BP algorithms have the shortcomings of converging slowly and learning unstably in feature-level fusion recognition. Based on the detailed analysis of momentum BP algorithm used in feature-level fusion recognition, a new momentum BP algorithm was presented, the resilient global-adaptive momentum BP algorithm （RGMOBP）. The algorithm backtracks the weights and reduces the learning rate when the previous weights-updating leads to an error increase. This weights-adjust method ensures that the error shift to the descent-orientation all the time. In the emulations, the RGMOBP algorithm gave much better results in learning performances than the other classical global-adaptive methods. They indicated that RGMOBP is an effective algorithm for feature level fusion recognition.