一种基于神经网络和遗传算法的拟人智能控制方法

A Human-imitating Control Based on Neural Networks & GA

提出一种基于Hopfield神经网络(HNN)和遗传算法(GA)混合策略的拟人智能控制方法。首先利用拟人智能控制得到定性控制律(线性或非线性),然后利用GA和HNN的混合优化策略实现定性控制律的定量化——首先,基于网格法产生GA的初始种群;然后,基于实数编码并采用最优个体保留策略、2/4择优选择以及引入控制经验的改进GA进行全局优化;最后,为了克服GA的后期收敛速度慢和局部优化能力缺乏,利用HNN的快速优化能力进行末段搜索,最终产生全局最优解。将该方法用于二级倒立摆系统的控制,仿真和试验结果均表明该方法有效。

A control method was proposed in this paper, which combines the human-imitating control theory and the optimization capability of Hopfield neural networks (HNN) and Genetic Algorithm (GA). Firstly, a qualitative control law (linear or nonlinear control law) is formed and relations among feedback gains are defined according to human-imitating control. The qualitative control law is then quantified by HNN & GA: at first, in order to increase the popularity and diversity of individuals, a grid method is used to produce the initial populations of the GA; then, real coding, elitism, 2/4 selection, arithmetic crossover, mutation, and human control experience are used in GA to generate the near global optimal solutions; finally, in order to improve the optimization speed and quality, HNN is used to execute the final optimization. Applying this method to control a double inverted pendulum system, the results of numerical simulations and experiments both demonstrate the validity of this control method.