基于反馈调控参数的BP学习算法研究

Research on back-propagation learning algorithm based on weight adjustment parameters

为解决经遗传算法优化后的BP网络极易陷入饱和区域而导致网络学习停滞的问题,基于神经生理解剖学关于神经电位脉冲发放系统和神经递质系统的耦合机理,提出一种改进的基于反馈调控参数的BP学习算法,通过反馈调控参数对神经元的节点输出进行扰动,避免学习过程中发生权值调整量趋于0的问题,从而解决经遗传算法优化后的BP网络容易出现的饱和区域问题.仿真实验结果表明,该方法能有效克服饱和区域引起的学习停滞问题,提高BP网络对遗传算法优化结果的精确定位能力,而且还具有收敛速度快和稳定性好的优点和在较大权值空间中的寻优能力.

When the initial weights and thresholds of back-propagation network are optimized by genetic algorithm, the learning process is more likely trapped into “flat spot” which results in slow or even suppressed learning process and weight adjustment due to the large scope of weight-initializing in genetic algorithm. To solve this problem and inspired by the research fruits of neuroscience and physiological anatomy, an improved learning algorithm based on weight adjustment parameters is proposed. It avoids premature saturation problem that the weight adjustment reaching zero by introducing an excitation unit into the neuron model to excitation the back-propagated error signal. Simulation results show that the new method can avoid the learning process being trapped into the “flat spot” efficiently and improve the capability of fine-tuning to the result optimized by the genetic algorithm. The new algorithm also outperforms the other traditional methods in terms of the convergence rate, stability, and the capability of searching optimum in large space.