An Iterative Approach to Automate the Tuning of Continuous Controller Parameters

Cyber-physical systems evolving in uncertain environment endure fluctuating dynamics during their lifetime. In such a variable context, controlling systems towards safety and system performances is challenging. In particular, controller tuning (finding optimal control parameters) is a challenging process due to the multiplicity of contexts to be considered. In this paper, we use a combination of model-driven simulation, dimensionality reduction, clustering and prediction techniques to define adequate control parameter settings. First, we propose to explore the controller behavior by simulating different configurations, a configuration is defined by a context (controlled process, environment, sensors, actuators) and a control parameters setting. From simulation results, a discretization is performed by binning the evaluation of quality of control. Then, we apply feature selection algorithms to identify contextual parameters that have a significant impact on performances of the controller. Considering only selected parameters, we finally carry out a clustering aiming at identifying for context domains an optimal control parameter setting. The approach is iterative to define the boundaries of the controller for a given context domain. For non simulated contexts, we propose a prediction module based on regression techniques.To evaluate the proposed approach, we compare it with classical control theory and we apply it to a proportional controller used for a leader/follower application. The experiment shows effectiveness in the identification of control parameters setting for different contexts.