基于遗传—模拟退火算法的过程神经网络的训练及应用

The Training and Application of Process Neural Network based on Genetic-Simulated Annealing Algorithm

过程神经网络是一种新型的神经网络,其输入及权值皆为时变函数,因此存在学习算法复杂度高、对初值敏感的问题。本文鉴于BP算法的不足提出了一种过程神经网络的学习算法,将输入函数和网络权值按正交基展开的过程神经网络,采用遗传算法与模拟退火算法相结合的方式进行网络训练,既避免了陷入局部最优解,又克服了模拟退火算法达到最优解造成的迭代次数增加问题,使网络具有较快的收敛速度和较高的逼近精度,文中给出相应的学习步骤和参数选取方法,同时以水淹层识别实验为例,验证了算法的有效性。

Process neural network （PNN） is a new neural network, whose inputs and weights are time-varying functions. While there exist some problems for PNN that the learning algorithm has high complexity and is sensitive to initial value. Considering the deficiency of BP algorithm, a learning algorithm of PNN is presented. PNN trains the network by combi- ning genetic algorithm and simulated annealing algorithm, whose input functions and network weights are expanded based on orthogonal basis. The algorithm of PNN avoids trapping into local minimum, and overcomes the problem that the iteration number of simulated annealing algorithm will be increased when finding the optimal solutions, which makes the network have fast convergence rate and high approximation accuracy. The corresponding learning steps and parameter selec- tion method are given in this paper. The experiment of water flooded layer identification is taken as an example to verify the effectiveness of the algorithm.