Self-Optimizing Control of Stochastic Systems

Optimal control of stochastic systems involves finding control strategies that optimize certain performance criteria while accounting for the parametric uncertainties and stochastic additive disturbances involved in the system dynamics. Model predictive control (MPC) solves an open-loop constrained stochastic optimal control problem repeatedly in a receding-horizon manner, resulting in large computation sometimes. Alternatively, the proposed stochastic self-optimizing control (SOC) selects optimal nonlinear controlled variables (CVs) offline by minimizing the expectation of the weighted closed-loop loss function based on neural network training. The nonlinear self-optimizing CVs are simply kept constant online so that the satisfactory control performance can be achieved. The proposed stochastic SOC requires much less online computation time compared with MPC, which is demonstrated by a two-mass spring simulation model.