Auto-tuning of controller parameters based on a probabilistic dynamic model with application in boat path following

Abstract

Tuning the parameters of controllers for intelligent boats poses a formidable challenge for engineers, demanding substantial time and effort. Traditionally, engineers rely heavily on their theoretical knowledge and professional expertise, coupled with trial-and-error approach, to fine-tune these settings. To alleviate this burden, this paper introduces an auto-tuning algorithm for boat path following controllers based on a probabilistic dynamic model. This algorithm adopts a model-based reinforcement learning paradigm, facilitating the autonomous calibration of controller parameters even with limited datasets. We use probabilistic ensemble method to construct the boat's probabilistic dynamic model which incorporates uncertainties in a principled manner, perform state propagation using particle-based methods, and optimize controller with particle swarm optimization. To validate the efficacy of our approach, we applied the proposed method to optimize two lateral controllers--a proportion-differentiation controller and a neural network controller--within the Gazebo simulation environment. The experimental outcomes underscore the high data efficiency of our method, demonstrating its ability to optimize controller parameters with minimal data requirements, thereby eliminating the need for manual expertise and time-intensive tuning procedures.