Multi-objective Longitudinal Decision-making for Autonomous Electric Vehicle: A Entropy-constrained Reinforcement Learning Approach

Abstract

The challenging task of “autonomous electric vehicle” opens up a new frontier to improving traffic, saving energy and reducing emission. However, many driving decision-making problems are characterized by multiple competing objectives whose relative importance is dynamic, and that makes developing high-performance decision-making system difficult. Therefore, this paper proposes a novel entropy-constrained reinforcement learning (RL) scheme for multi-objective longitudinal decision-making of autonomous electric vehicle. Firstly, in order to prevent the policy from prematurely converging to a local optimum, the policy’s entropy is embedded in proximal policy optimization (PPO) algorithm based on actor-critic architecture. Secondly, a self-adjusting mechanism to the weight of entropy is developed to accelerate model training and improve algorithm stability through entropy constraint. Thirdly, multimodal reward signals are designed to guide the RL agent learning complex multi-modal driving policies by considering safety, comfort, economy and transport efficiency. Finally, simulation results show that, the proposed longitudinal decision-making approach for autonomous electric vehicle is feasible and effective.