Constrained Bayesian Optimization of a Linear Feed-Forward Controller

Abstract

Models of dynamic systems often contain uncertain parameters or do not describe all dynamics. In some cases, they cannot represent the plants accurately enough, which limits the achievable performance of model-based controller design. Learning-based controllers can adapt to the true parameters and cope with unmodeled dynamics. However, it must be ensured that an adaption does not cause the system to violate its physical or operating constraints, which could result in dangerous situations or harm the environment. We consider the case of a repetitive executed trajectory and propose a learning feed-forward controller that safely improves the tracking performance from iteration to iteration without violating predefined constraints. Therefore we construct a FIR filter whose parameters are optimized through Bayesian Optimization with constraints. We demonstrate the performance and compliance with the constraints during the learning process on a linear system with actuator dynamics. Furthermore, we show that the approach outperforms norm optimal iterative learning control.