Advanced multivariable controllers are widely used in process and other industries. However, such controllers require regular maintenance to sustain acceptable closed loop performance. It is common practice to monitor controller performance and intermittently initiate a model re-identification procedure in the event of performance degradation. However, such procedures are complicated and resource-intensive, and they often cause costly interruptions to normal operations.

 

I will present several recent results on exploiting Deep and Meta Reinforcement Learning (RL) for automatic maintenance of controllers. I will start the presentation with simple fixed-structure linear controllers and use RL as a supervisory mechanism to maintain the controllers. I then extend this approach to develop an adaptive controller using Deep and Meta RL. I will highlight and offer solutions to practical aspects of these controllers such as closed-loop stability and sample efficiency.

 

This work was done in collaboration with Nathan Lawrence, Daniel McClement, Philip Loewen, Michael Forbes, and Johan Backstrom.