Inverse Reinforcement Learning Based Driver Behavior Analysis and Fuel Economy Assessment
Abstract
Human drivers have different driver behaviors when operating vehicles. These driving behaviors, including the driver’s preferred speed and rate of acceleration, impose a major impact on vehicle fuel consumption consequently. In this study, we proposed a feature-based driver behavior learning model from demonstrated driving data utilizing the Inverse Reinforcement Learning (IRL) approach to analyze various driver behaviors and their impacts on vehicle fuel consumption. The proposed approach models the individual driving style as cost function which is a linear combination of the features and their corresponding weights. The proposed IRL framework is used to find the model parameters that fit the observed driving style best. By using the learned driving behavior model, the most likely trajectories are computed and the optimized acceleration feature weights are used as driving style indicators to analyze different driver behaviors. The different driver behaviors and their impacts on vehicle fuel consumption are then analyzed with different drivers in real-world driving scenarios. Results show that the proposed IRL framework can successfully learn individual driver behaviors using vehicle trajectory data demonstrated by different real drivers. The learned driver behaviors promise a significant correlation between driving behavior and fuel consumption.
