Adaptive Optimization Control Of Nonlinear Systems Based On Deep Reinforcement Learning

In this paper,an adaptive optimal controller design algorithm based on deep reinforcement learning algorithm is studied to solve the problem of adaptive optimal control for continuous time nonlinear systems.For the nonlinear systems with complex or unknown models,it is difficult to design the optimal control algorithm from the perspective of the model due to the complexity and unpredictability of the system model.Considering the decision making ability of reinforcement learning and the environmental perception ability of deep learning.In this paper,three adaptive optimization controllers for continuous time nonlinear systems are proposed.The biggest advantage of the deep reinforcement learning algorithm proposed in this paper is that it makes full use of the latest deep learning technology and combines deep learning with reinforcement learning to solve the adaptive optimization controller of the complex and unknown continuous time nonlinear system.The main work and contributions of this paper are embodied in the following:First,a new online adaptive optimal controller is proposed for a class of nonlinear continuous time systems with input delay,whose model is partially unknown.The linear differential inclusion technique is used to linearize the original system,and the adaptive optimal controller of the linearized system is obtained through the online strategy iterative algorithm,and the convergence of the adaptive optimal control algorithm is proved.Finally,the effectiveness of the proposed method is verified by two simulation examples.Then,the design of adaptive optimal controller for a class of nonlinear systems with unknown continuous time model is studied.Combining Q-learning algorithm and generative adversarial network scheme,a new adaptive optimal control algorithm for unknown nonlinear systems with continuous time model is successfully designed.The new generation adversarial network training strategy is adopted to stabilize the system and the convergence of the proposed adaptive optimal control algorithm is proved.Finally,the effectiveness of the proposed method is verified by a simulation example,and the superiority of the proposed algorithm is illustrated by comparing with the traditional role-critic algorithm.Next,considering the optimal control process of most complex industrial systems,it is difficult,time-consuming and expensive to determine an accurate mathematical model of the cost function.To solve this problem,a deep element reinforcement learning algorithm based on cost prediction is proposed to solve the optimal controller.Using the latest codec structure to construct the cost function network and combining the meta-learning algorithm and the reinforcement learning scheme,an optimal control design method which can adapt to different actual task environments is successfully designed.Finally,a simulation example is given to demonstrate the effectiveness and superiority of the proposed method.Finally,it gives a summary and prospect,and points out the problems to be further solved and improved.