Efficient Task Assignment in Dependency-Cooperative Spatial Crowdsourcing

Abstract

With the rapid advancement of mobile technology and ubiquitous computing, spatial crowdsourcing has emerged as a promising computing paradigm. The growing complexity of spatial tasks has increasingly driven the demand for coordination among interdependent tasks and the cooperative participation of mobile workers, which remains relatively unexplored among existing studies. To fill this gap, this paper introduces Dependency-Cooperative Spatial Crowdsourcing (DCSC), a novel problem that explicitly models task dependencies and enables systematic worker cooperation to address complex spatial scenarios. To solve DCSC, we propose a two-stage solution: Dependency-aware Recommendation and Cooperation-aware Matching. In the first stage, we develop a multi-agent reinforcement learning approach enhanced with meta-gradient techniques to recommend suitable subtasks while considering dependency constraints. In the second stage, we propose a genetic algorithm-enhanced game approach to achieve optimal cooperative assignment, guided by a multi-dimensional matching utility function. To ensure consistency and optimization across both stages, we employ meta-gradients from the policy network to guide the optimization of the utility function. Additionally, we utilize graph neural network-based policy clustering to address task heterogeneity, enabling each cluster to learn specialized parameters and enhance reinforcement learning performance. Extensive experiments validate the effectiveness of our approach.