Explainable AI methods on a deep reinforcement learning agent for automatic docking

摘要

Artifical neural networks (ANNs) have made their way into marine robotics in the last years, where they are used in control and perception systems, to name a few examples. At the same time, the black-box nature of ANNs is responsible for key challenges related to interpretability and trustworthiness, which need to be addressed if ANNs are to be deployed safely in real-life operations. In this paper, we implement three XAI methods to provide explanations to the decisions made by a deep reinforcement learning agent: Kernel SHAP, LIME and Linear Model Trees (LMTs). The agent was trained via Proximal Policy Optimization (PPO) to perform automatic docking on a fully-actuated vessel. We discuss the properties and suitability of the three methods, and juxtapose them with important attributes of the docking agent to provide context to the explanations.