A New Solution to the PID18 Challenge: Reinforcement-Learning-based PI Control

The PID18 benchmark challenge, first presented at the 3rd IFAC Conference on Advances in 2018, is widely used to test the effectiveness of proportional-integral-derivative (PID) controllers. In this challenge, a 2-Input and 2Output vapor-compression refrigeration system with strong nonlinearities and couplings is used as the test control plant. In this paper, we address this challenge by proposing a new reinforcement-learning-based PI controller. Unlike current existing results, the PI parameters of each RL-based PI controller are determined by two adaptive signals, which are generated by a reinforcement learning (RL) agent, aiming at achieving better control performance. Extensive comparisons with existing methods have been conducted, and it is shown that our proposed strategy yields the best control performance thus far.