RNN和Hopfield两网络中的优化解法的对比研究

Comparison Study of Optimization Solution between Random Neural Network and Hopfield Network

基于一种动态随机神经网络 (DRNN)求解典型NP优化问题TSP的改进算法 ,在理论上对DRNN与连续的Hopfiled网络(CHNN)进行了对比研究 ,指出虽然两种网络均以能量函数表达TSP的最优路径 ,并通过训练反馈网络求得路径解 ,但由于两者所用激活函数和收敛条件不同 ,使得DRNN网络能够接受能量函数的小波动 ,从而跳出局部最小值达到全局最优 ;此外 ,DRNN与CHNN相比网络训练对参数变化不敏感 ,参数设置简单。最后 ,通过仿真实验对随机坐标十城市使用两种网络对比路径寻优能力 ,进一步验证理论分析的结论。

Based on the improved algorithm of Dynamic Random Neural Network (DRNN) on the typical NP problem - TSP, the comparison theoretical study of DRNN and Continue Hopfield Neural Network (CHNN) is analyzed. The two networks both use energy function as the expression of the final path solution by training the feedback networks, but the difference of working rules, convergence conditions makes DRNN accept small fluctuation of the energy function to escape the local minimum and reach the global one. On the other hand, compared with CHNN, the training parameters of DRNN are less sensitive and easier to settle. The theoretical conclusions are validated by experiments of two networks on the 10-city TSP coordinating randomly. The advantages and disadvantages of the two networks are discussed.