一种基于改进遗传算法的神经网络优化算法研究

遗传算法是目前优化搜索算法中应用比较广泛的一种,但基本遗传算法存在收敛速度慢、易于陷入局部最优等缺点。针对上述问题对遗传算法(GA)的选择算子进行改进,在最优保存策略的基础上将每代种群按照适应度由小到大排序,平均分成前中后3段,按照0.6、0.8、1的比例进行选择;从尾段中随机抽取个体来补足种群由于选择操作而损失的个体;既利用了最优保存策略的全局收敛特性同时也保持了种群的多样性;用改进的遗传算法调整神经网络的权值形成了新的改进遗传算法优化BP神经网络(IGA-BP);通过与选择算子为适应度比例选择算子的GA-BP网络进行比较,结果表明算法改进后缩短了收敛时间同时减少了运行误差;最后将该改进算法应用于水泥回转窑的故障诊断中,验证了算法的可行性。

Study on a neural network optimization algorithm based on improved genetic algorithm

Genetic algorithm is one of the more widely used in the optimization search algorithms at present; however, basic genetic algorithm has the defects of slow convergence speed and easy to fall into local optimum. To solve these problems, it is proposed to improve the selection operator of genetic algorithm (GA). On the basis of the elitist strategy, the population of each generation is sorted according to the fitness in ascending order, and then is uniformly divided into three sections, i.e. front, middle and tail sections; in the three sections the numbers of individuals are selected according to the proportions of the original numbers of individuals of 0.6, 0.8 and 1, respectively. The individuals randomly selected from the tail section of the population are used to replenish the lost individuals due to the selection operation. This will not only take the advantage of the global convergence property of elitist strategy, but also maintain the diversity of the population. Through using the improved genetic algorithm to adjust the weights of the neural network, a new improved genetic algorithm optimization BP neural network (IGA BP) is formed. The improved algorithm is compared with the GA BP network whose selection operator is fitness ratio selection operator; the results show that the improved algorithm shortens the convergence time and reduces the operation error. Finally, the improved algorithm was applied to the fault diagnosis of a cement rotary kiln, the result verifies the feasibility of the improved algorithm.