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

Abstract

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