Sequential Learning for Adaptive Critic Design: An Industrial Control Application

This paper investigates the feasibility of applying reinforcement learning (RL) concepts to industrial process optimisation. A model-free action-dependent adaptive critic design (ADAC), coupled with sequential learning neural network training, is proposed as an online RL strategy suitable for both modelling and controller optimisation. The proposed strategy is evaluated on data from an industrial grinding process used in the manufacture of disk drives. Comparison with a proprietary control system shows that the proposed RL technique is able to achieve comparable performance without any manual intervention