Multi-objective optimization for autonomous driving strategy based on Deep Q Network

Abstract

Autonomous driving is an important development direction of automobile technology, and driving strategy is the core of the autonomous driving system. Most works in this area focus on single-objective tasks, such as maximizing vehicle speed or lane-keeping, and rare attention has been paid to the quality of driving skills. Therefore, a multi-objective learning method is proposed for autonomous driving strategy based on deep Q-network, where two optimization objectives are involved, i.e., vehicle speed and passenger comfort. An end-to-end autonomous driving model is designed by using vehicle front camera images as inputs to the Q-network and makes decisions based on the output Q values. Considering the vehicle speed and passenger comfort, the reward function is designed for multi-objective optimization. To evaluate the effectiveness of the method, training and testing are performed in a simulator, and a single-objective strategy with the goal of maximizing speed is designed for comparison. The results show that the proposed multi-objective autonomous driving strategy can strike a balance between vehicle speed and passenger comfort. Compared with the single-objective strategy, the multi-objective strategy has a significant improvement in comfort, while the average speed is only slightly reduced.