A Dual-Embedding Based Reinforcement Learning Scheme for Task Assignment Problem in Spatial Crowdsourcing

Yucen Gao, Dejun Kong, Haipeng Dai, Xiaofeng Gao, Jiaqi Zheng, Fan Wu, Guihai Chen

Abstract: With the popularity of mobile devices, spatial crowdsourcing has attracted widespread attention, which collects spatial tasks with location constraints and assigns them to workers who can travel to certain locations to participate in and obtain profits. One of the core issues is task assignment, in which tasks should be assigned to proper workers to maximize the overall utilities. In the paper, we consider a Utility-driven Destination-aware Spatial Task Assignment (UDSTA) problem, where the utility of a worker is modeled as the completed task profit minus the worker’s travel cost, which is more realistic and involves route planning while assigning tasks. We prove that this problem is NP-complete and propose a dual-embedding based deep Q-Network (DE-DQN) to sequentially assign tasks to proper workers. Specifically, we design a utility embedding to reflect the top-k utility tasks for workers and worker-task pairs, and a coverage embedding to represent the potential future utility of an assignment action. The state of DQN consists of the utility embedding, remaining workload, and cumulative utility. Besides, the action of this DQN is formed by concatenating the utility and coverage embedding. We also provide an enhanced version called DE-Rainbow by using Rainbow DQN instead of traditional DQN for further optimization. For the first time, we combine the dual embedding with DQN to achieve a multi-task and multi-worker matching and obtain the route plans of workers. Experiments based on both synthetic and real-world datasets indicate that DE-DQN and DE-Rainbow perform well and show significant advantages over the baseline methods.

External IDs:dblp:journals/www/GaoKDGZWC25