Ensemble Reinforcement Learning-Based Supervisory Control of Hybrid Electric Vehicle for Fuel Economy Improvement

This study proposes an ensemble reinforcement learning (RL) strategy to improve the fuel economy. A parallel hybrid electric vehicle model is first presented, followed by an introduction of ensemble RL strategy. The base RL algorithm is Q -learning, which is used to form multiple agents with different state combinations. Two common energy management strategies, namely, thermostatic strategy and equivalent consumption minimization strategy, are used as two single agents in the proposed ensemble agents. During the learning process, multiple RL agents make an action decision jointly by taking a weighted average. After each driving cycle iteration, Q -learning agents update their state-action values. A single RL agent is used as a reference for the proposed strategy. The results show that the fuel economy of the proposed ensemble strategy is 3.2% higher than that of the best single agent.