Virtual Platform for Reinforcement Learning Research for Heavy Vehicles

Abstract

The main objective of the project, tasked by Volvo Group Trucks Technology, is to implement a reinforcement learning agent in a open-source vehicle simulation model known as VehProp, developed at Chalmers University of Technology in MathWorks Simulink. The project also aims to improve VehProp, particularly the Equivalent Consumption Minimization Strategy. As a proof of concept for the reinforcement learning implementation, an agent is trained to control the brakes of a hybrid electric heavy duty vehicle in order to minimize fuel consumption through regenerative braking. Much effort is put in the theory chapter, explaining reinforcement learning and the simulation platform. The reinforcement learning agent is successfully implemented in the simulation platform, using the Reinforcement Learning Toolbox in Matlab. The training of the agent to control the brakes of a hybrid electric heavy duty vehicle was unsuccessful, with the agent failing to display the wanted behavior. Suggestions for future work with the agent are presented, mainly fulfilling the Markov property, investigating sparse reward functions and the general implementation of the agent to assure action-reward causality. Improvements to the Equivalent Consumption Minimization Strategy of the simulation platform were made with a decrease in fuel consumption as a result.