Abstract: Aiming at the problem that most of the existing task assignment in spatial crowdsourcing only considered unilateral benefits, short-term benefits and single scenario, we proposed a spatial crowdsourcing task assignment algorithm based on multi-agent deep reinforcement learning. Firstly, a new spatial crowdsourcing scenario was defined, in which workers could freely choose whether to cooperate with others. Secondly, a multi-agent deep reinforcement learning model based on the attention mechanism and A2C (advantage actor-critic) method was designed for task assignment in the new scenario. Finally, simulation experiments were carried out, and the performance of the algorithm was compared with other latest task assignment algorithms. The experimental results show that the proposed algorithm can achieve higher task completion rate and worker profitability rate simultaneously, which proves the effectiveness and robustness of the algorithm.

Key words: multi-agent deep reinforcement learning, spatial crowdsourcing, task assignment, attention mechanism