Freshness aware vehicular crowdsensing with multi-agent reinforcement learning

Abstract

Vehicular crowdsensing leverages the mobility and sensing capabilities of vehicles to provide efficient data collection and monitoring services for urban areas. However, maintaining data freshness in urban sensing environments while addressing issues such as complex spatiotemporal data correlations, dynamic city conditions, and the trade-off between task performance and costs remains a significant challenge. To address this issue, we propose freshness aware Multi-Vehicular Crowdsensing (freshMVCS), a decentralized multi-agent deep reinforcement learning framework for long-term vehicular scheduling in data collection tasks. Following the decentralized training decentralized execution paradigm, each agent in freshMVCS is embedded with an independent recurrent neural network and intrinsic reward mechanism to enhance exploration capabilities, while achieving collaboration through shared task information. Extensive experiments conducted on real-world datasets demonstrate that the freshMVCS approach excels in maintaining data freshness, achieving high collection rates, and minimizing Age of Information threshold violations. These results indicate its effectiveness in accomplishing long-term data collection tasks within complex and dynamic urban sensing environments.