Safe Active Learning and Safe Bayesian Optimization for Tuning a PI-Controller

Abstract

In the combustion engine calibration domain, many controllers are still tuned manually or using simple adjustment laws. In order to increase workforce efficiency, automated methods for controller tuning are desirable. Often, the structure of the controller is fixed and only its parameters have to be optimized. Model based controller tuning methods require a good dynamic model of the system. Such models are often hard to obtain, as deep system knowledge or extensive measurements at the system, potentially in open loop, may be required. In some cases only closed loop measurements are possible, for example, due to system instability. If controller tuning methods interact with the real system for which the controller shall be tuned, they have to comply with safety constraints. For example, parameter sets resulting in unstable control loops or ones causing critical system states as with very high overshoot should not be tested at the real system.In this contribution, two optimization-based methods for tuning controller parameters are compared. The first method, Safe Active Learning for control, learns a loss function based on controller parameters. Subsequently, an offline optimization is pursued. The second method, a newly proposed Safe Bayesian Optimization algorithm, combines learning of a loss function model with online optimization. Both methods perform closed loop measurements and take safety constraints into account. The methods are evaluated and compared at a PI controller of a real high pressure fuel supply system in a test vehicle.