Chance constrained policy optimization for process control and optimization

Chemical process optimization and control are affected by (1) plant-model mismatch, (2) disturbances, (3) constraints for safe operation. Reinforcement learning policy would be a natural way to solve this due its ability address stochasticity, directly account the effect of future uncertainty feedback in proper closed-loop manner; all without need an inner loop. One main reasons why reinforcement has not been considered industrial processes (or almost any engineering application) is that it lacks framework deal with safety critical constraints. Present algorithms use difficult-to-tune penalty parameters, fail reliably satisfy state or present guarantees only expectation. We propose chance constrained (CCPO) algorithm which satisfaction joint high probability — crucial tasks. This achieved introduction constraint tightening (backoffs), computed simultaneously policy. Backoffs adjusted Bayesian using empirical cumulative distribution function probabilistic constraints, therefore self-tuned. results general methodology can imbued into enable them probability. case studies analyse performance proposed approach. • Chance technique constructed. have combined produce use. Case parametric structural considered. Offline optimal used online.