Explainable Reinforcement Learning Models for Transparent Autonomous Vehicle Control

Abstract

[1] [2]The development of autonomous driving technologies has generated broad research interests and become a matter of societal importance. Learning algorithms, including reinforcement learning, have an important role to play in such safety-sensitive applications. A reliable autonomy solution should be capable of explaining its reasoning and decisions to users in a transparent and comprehensible manner. This has motivated an increasing effort towards the development of explainable AI and trustworthy autonomy solutions. In this context, this section highlights a main open challenge in the development of reinforcement learning (RL) agents for traffic control and autonomous vehicle (AV) decision making: learning explainable models aiming at human-friendly AV decision making.[3]The main application of RL in AVs is in decision making, including high-level navigation policy and low-level robot control. In this domain, the interaction between the model and the simulated or real environment makes it difficult to easily understand how the model “sees” certain situations, drives the car, recognizes drives, or reacts to pedestrians.