【摘要】 在马尔可夫决策模型框架下,提出一种基于轨迹分析的计算评估方法,通过分析驾驶回报设置和车辆轨迹的特征期望衡量自主导航系统的性能。假定回报函数是回报特征的线性组合,通过逼近不同的车辆自主驾驶策略,求解应用于沙盒场景的回报设置,从而仿真导航轨迹的特征期望。实验结果表明,该方法能对自主导航系统的轨迹数据实现定性和定量评估。更多还原

【Abstract】 This paper proposes a novel trajectory-analysis-based performance evaluation method,which assesses and compares the navigation performances by reward settings and feature expectations under Markov Decision Process(MDP) framework.Assuming that the reward function is a linear combination of known features,this method evaluates reward settings by approximating the driving styles of different land autonomous navigation systems.On this basis,reward settings are used to simulate the feature expectations of navigation trajectories in a standard land navigation environment called the sandbox.Experimental results show the validity of this method in both qualitative and quantitative performance evaluations.更多还原