Research on Deep Reinforcement Learning of Unmanned Multi-sensor Fusion Based on Adaptive Reward Mechanism

Based on the problems of low learning efficiency, poor stability and weak generalization ability in the unmanned driving decision-making algorithm, this paper proposes two solutions: For the problem of low generalization ability caused by single sensor input data, this paper propose a multi-sensor fusion deep reinforcement learning algorithm, which reduces information redundancy through differential design and realizes the feature complementarity of each sensor. For the problems of the traditional reward function design method, this paper design a Multi-objective adaptive Markov Decision Processes reward function (MAMDP), which realizes the adaptive and real-time dynamic evaluation of the decision system. The experimental results show that the convergence, efficiency and stability of this Adaptive reward Deep Deterministic Policy Gradient algorithm (ADDPG) have been significantly improved, which effectively improves the accuracy and stability of driverless decision-making behavior.