Proposal of Hybrid Controller Based on Reinforcement Learning for Temperature System

Abstract

The ability of a modern controller in addressing fast convergence of existing adaptive PID controllers has a restriction. This case applies to several actuators, including thermostat system. In this project, Reinforcement Learning (RL) is applied to the challenge in the automatic tuning of a proportional-integral-derivative controller. Two means of testing procedures will be carried out in this project to achieve the objectives at investigating the unprecedented performance of the hybrid controller. Firstly, the researchers combine the asynchronous learning structure of the Asynchronous Advantage Actor-Critic (A3C) with the incremental PID controller. Secondly, the researchers also unite a PID system with a deep deterministic policy gradient (DDPG). Both actor network structure and critic network structure are used back-propagation neural networks with three-layer structure. A comprehensive review of the literature relating to the hybrid controller is also provided.