A Deep Reinforcement Learning Approach to Eco-driving of Autonomous Vehicles Crossing a Signalized Intersection

Abstract

: This paper outlines a method for obtaining the optimal control policy for an autonomous vehicle approaching a traffic signal head. It is assumed that traffic signal phase and timing information can be made available to the autonomous vehicle as the vehicle approaches the traffic signal. Constraints on the vehicle’s speed and acceleration are considered and a microscopic fuel consumption model is considered. The objective is to minimize a weighted sum of the travel time and the fuel consumption. The problem is solved using the Deep Deterministic Policy Gradient algorithm under the reinforcement learning framework. First, the vehicle model, system constraints, and fuel consumption model are translated to the reinforcement learning framework, and the reward function is designed to guide the agent away from the system constraints and towards the optimum as defined by the objective function. The agent is then trained for different relative weights on the travel time and the fuel consumption, and the results are presented. Several considerations for deploying such reinforcement learning-based agents are also discussed.