Performance-oriented model learning for control via multi-objective Bayesian optimization

• The problem of performance-oriented model learning for model-based control is formulated as a black-box, multi-objective optimization problem. solved using Bayesian optimization, which particularly suitable handling expensive and noisy performance measures in closed-loop control. Performance-oriented adaptation can yield control-relevant models that result significant improvement compared to identification data. controllers, such predictive control, largely depends on the quality their underlying system dynamics. Inspired by notion this paper presents strategy performance-oriented, data-driven fundamental idea mitigate plant-model mismatch via improving model’s control-oriented optimizing interest, opposed enhancing its general accuracy. To end, we leverage composite structure consists prior (physics-based or data-driven) be efficiently adapted manner towards maximization performance. solve problem, use (MOBO) directly handle black-box expensive-to-evaluate functions are computed from observations measures. MOBO approach demonstrated benchmark bioreactor case study. Simulation results indicate that, given fixed budget process runs, performance, data, while systematically accounting multiple proposed useful auto-tuning controllers processes with finite-time objectives, where each run associated high monetary value.