关键词: 共享单车调度, 需求预测, 用户激励, 马尔科夫决策, 深度强化学习

Abstract: Focused on the issue that bike-sharing dispatching, considered budget constraints, restrictions on users' maximum walking distance, users' temporal and spatial needs, and dynamic changes in the distribution of shared bicycles. Devised a bike-sharing dispatching strategy with user participation to improve the long-term user service rate of the platform. The dispatching strategy includes task generation algorithm, budget allocation algorithm and task allocation algorithm. In the task generation algorithm, predicted the user’s future bicycle demand based on LSTM. In the budget allocation algorithm, the sequential allocation of budgets for each time period is a sequential decision-making problem, and thus modeled it as a Markov decision process. At the same time, considering that the problem has a high-dimensional and continuous state space and a continuous action space, designed a budget allocation strategy based on the deep deterministic strategy gradient algorithm DDPG. In the task allocation algorithm, due to the budget constraint that makes it impossible to use the mainstream bipartite graph matching algorithm, used the greedy matching strategy for task allocation. Finally, we run experiments based on the Mobike dataset to evaluate our strategy against the dispatching strategy with unlimited budget, the dispatching strategy with greedy budget allocation, the dispatching strategy under truck hauling, and the situation without dispatching. The results show that our shared bicycle dispatching strategy with user participation can achieve the best results except for the dispatching strategy with unlimited budget. The experimental results can provide some useful insights for dispatching shared bikes.

Key words: Bike-Sharing Dispatching, Demand Prediction, User Incentive, Markov decision, Deep Reinforcement Learning