Parameter Optimization of Adaptive PID Controller Based on Deep Reinforcement Learning

This paper proposes a deep reinforcement learning-based adaptive PID controller. The strategy in reinforcement learning provides adaptability. Real-time optimizations of control parameters are performed in a PID controller. Traditional PID controllers face the problem of insufficient adjustment of parameters while dealing with uncertain and nonlinear systems. In this work, the gain parameters adjustment of the PID is performed by a deep Q-network algorithm so that it rapidly adapts to different environmental systems' requirements. Experiments test the response of each control method under conditions of no uncertainty, slight uncertainty, and high uncertainty, and compare the system response curve under parameter mutation scenarios. Experimental results show that the DRL-PID controller exhibits obvious superior robustness and stability in handling model uncertainty and parameter jumps, effectively reducing overshoot while speeding up the response time. Compared with a conventional PID controller and GA-based PID controller, the proposed DRL-PID controller exhibits much better control performance under various kinds of uncertain environments. The results in the present study provide a novel approach to the intelligent control for complex dynamic systems and applications to a wide range.