关键词: 物资分配, 强化学习, Q-learning算法, 动态探索率, 动态Boltzmann Softmax

Abstract: Aiming at the problem of emergency resource allocation for natural disasters, we proposed a Q-learning algorithm based on dynamic Boltzmann Softmax (DBS) and dynamic exploration rate (DER) (DBSDER-QL).  Firstly, the DBS strategy was used to dynamically adjust the weights of action values, promoting stable convergence of the algorithm and solving the problem of excessive of  the maximum operator. Secondly, the DER strategy was used to improve convergency and stability of the algorithm, solving the problem of the fixed exploration rate Q-learning algorithm not fully converging to the optimal strategy in the later stage of  training. Finally,  the effectiveness of the DBS and DER strategies was verified by ablation experiments. Compared with 
 dynamic programming, the greedy algorithm, and traditional Q-learning algorithm, the experimental results show  that DBSDER-QL algorithm is significantly better than traditional methods in terms of total cost and computational efficiency, showing higher applicability and effectiveness.

Key words: resource allocation, reinforcement learning, Q-learning algorithm, dynamic exploration rate, dynamic Boltzmann Softmax