Adaptive decision making is critical for safety and comfort of autonomous vehicles (AVs) driving in mixed traffic situations. However, existed decision-making algorithms focus on learning directly from expert behaviors while ignore the temporal efficiency and adaptiveness to environment changes, decreasing the flexibility of AVs. Therefore, this paper proposed a deep reinforcement learning (DRL) framework with temporal attention mechanism. First, we incorporate imitation strategies to learn from expert demonstrations to overcome the initial strategy search inefficiency in DRL, thus capture the implicit decision-making mechanisms that hidden in expert behavior data. Second, a temporal attention mechanism is embeded into the DRL framework, while utilizes an extended sequence of historical observations to analyze the temporary driving styles of other traffic participants, further facilitating smoother lane change operations of AVs. Last but not the least, we model high-level lane change decisions as composition of discrete sub-operations of acceleration and angular velocity of steering. This strategy enhances the control precision of AVs, enabling rapid and secure lane-change actions. Thousands of experiments were conducted in the CARLA simulator and the results show that our algorithm significantly outperform the state-of-the-art DRL algorithms and behavioral planners in multiple lane change scenarios.


    Access

    Check access

    Check availability in my library

    Order at Subito €


    Export, share and cite



    Title :

    TAIL-DRL:Adaptive Lane Change Decision Making Of Autonomous Vehicles In Mixed Traffic


    Contributors:
    Yu, Junru (author) / Xu, Risheng (author) / Wang, Xiao (author) / Mu, Chaoxu (author)


    Publication date :

    2024-07-05


    Size :

    259736 byte




    Type of media :

    Conference paper


    Type of material :

    Electronic Resource


    Language :

    English