An adaptive critic based neurocontroller for process control

This work describes the use of an online neurocontroller based on reinforcement learning for a process control problem. The approach used is an extension of the model-free action-dependent adaptive critic design (ACD) proposed by Si et al. (2001). This is used to develop an online learning strategy to optimise a nonlinear PI controller in the absence of an analytical plant model. Results are presented for a validated simulation of a laboratory temperature process. Comparison with conventional PI control shows that the technique is able to tune a nonlinear PI controller, without any manual intervention, and achieved comparable performance.