Destination-Oriented Data Collection and Uploading for Mobile Crowdsensing With Multiagent Deep Reinforcement Learning

Data collection and uploading are among the most fundamental problems in an mobile crowdsensing (MCS) system and involve complicated interactions among the system's platform, mobile users, data collection tasks, and data relaying devices. In the article, we propose a destination-oriented data collection and uploading (DDCU) problem in MCS to facilitate cooperation among various entities. In DDCU, mobile device users are recruited as workers to complete data collection tasks and aim to reach their destinations after collecting the required data. At the same time, edge servers are recruited as data relaying devices to share their idle network resources and gain extra profit. The DDCU problem aims to help an MCS platform plan efficient sensing paths for workers to maximize the platform's total profit. We further prove the DDCU problem to be NP-hard and thus require a time-efficient approximation-based method to solve the problem. In Xu and Song (2022), we proposed communication-QMIX-based multiagent deep reinforcement learning (CQDRL) as a decentralized method for an MCS routing problem. Although CQDRL was proven effective with high performance, it explicitly targets graph-based problems with two types of nodes and requires considerable effort to be extended to more heterogeneous graph problems. In this article, we propose CQDRL-S by simplifying the graph processing part of CQDRL to solve the DDCU problem with three types of nodes. More importantly, CQDRL-S can potentially be extended to other heterogeneous graph problems with moderate modifications. Extensive experiments further justify the effectiveness, efficiency, and simplicity of CQDRL-S in dealing with the DDCU problem.