Stochastic predictive control with adaptive model maintenance

Abstract

Abstract-The closed-loop performance of model-based controllers often degrades over time due to increased model uncertainty. Some form of model maintenance must be performed to regularly adapt the system model using closed-loop data. This paper addresses the problem of control-oriented model adaptation in the context of predictive control of stochastic linear systems. A stochastic predictive control approach is presented that integrates stochastic optimal control with control-oriented input design in order to confer some degree of probing effect to the control inputs. The probing effect will enable generating informative closed-loop data for (online) control-oriented model maintenance. In a simulation study, the performance of the proposed stochastic predictive control approach with integrated input design is demonstrated on a atmospheric-pressure plasma jet with potential biomedical applications.

Model predictive control with active learning under model uncertainty: Why, when, and how

“…Broadly, implicit dual control involves obtaining an approximate solution to the Bellman equation (), which has the learning component implicitly included through the hyperstate as discussed above . Explicit dual control, conversely, generally refers to replacing Problem 1 with a surrogate optimization problem that accounts for a measure of reducible model uncertainty . Thus, explicit methods incorporate some form of probing into the optimal control problem explicitly through approximation of Problem 1.…”

Information-theoretic multi-time-scale partially observable systems with relevance to leukemia treatment

“…We dis-cuss some recent papers on dual control, which involve systems with unknown parameters (exception: [11]). Model predictive controllers for linear systems with process noise [3] and non-linear systems without process noise [23] have been developed, where the objective is a sum of a performance metric and an FIM-based metric. The authors of [23] also assume initial state distributions with bounded support.…”

Stochastic Model Predictive Control: An Overview and Perspectives for Future Research

“…» Adaptive (dual) control: MPC integrated with persistent excitation has recently gained interest with the aim of ensuring uniform quality of system models in MPC applications [125], [126]. Combining SMPC with input design to consistently adapt model structure and/or parameters in the face of stochastic system uncertainties remains an open research challenge that lends itself to several theoretical issues (see [127] and [148]for recent results on SMPC integrated with input design). » Explicit stochastic predictive control: The benefits of explicit MPC include efficient online computations and verifiability of the control policy.…”