Intelligent task allocation and data uploading for Mobile Crowd Sensing
Loading...
Date
2023-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of New Brunswick
Abstract
Mobile crowd sensing (MCS) is a large-scale sensing paradigm based on existing communication infrastructures and crowd intelligence. MCS is an economical alternative to traditional sensor networks for collecting human-related data in urban areas with low deployment fees. In a practical MCS system, task allocation and data uploading are its main functional components and directly determine the system’s practicality on many occasions. A typical problem in task allocation or data uploading usually involves various entities, such as mobile devices, data collection tasks, MCS platforms and edge nodes. An effective MCS scheme should satisfy complex requirements from the entities and coordinate them to achieve collective sensing and profit goals.
This thesis focuses on task allocation and data uploading under different circumstances in MCS. Due to complex constraints and objectives, the problems in task allocation and data uploading are NP-hard. Thus, optimal solutions to the problems are usually impractical due to explosively increasing running time with growing problem scales, and heuristic methods are considered a practical alternative. On the one hand, heuristic methods that are handcrafted according to human institutions can generate time-efficient results. However, such handcrafted heuristic methods may experience inconsistent performance in some unexpected cases owing to the complex dynamics of NP-hard problems. On the other hand, metaheuristic methods can generate consistent performance with sufficient exploration in the solution spaces of problems. Nevertheless, intensive exploration can be time-consuming, and such metaheuristic methods are unsuitable in some time-sensitive scenarios.
This thesis proposes a series of deep reinforcement learning (DRL) based methods for NP-hard task allocation and data uploading problems to avoid the drawbacks of handcrafted and metaheuristic methods. Specifically, DRL-based methods memorize their experience over a group of problem instances into deep neural networks (DNNs) and then apply the experience to solve other problem instances. Thus, the DRL-based methods not only generate time-efficient results but also automatically adapt to various cases. Extensive experiments show that DRL-based methods can outperform baseline methods in various aspects.