An incentive mechanism for joint sensing and communication Vehicular Crowdsensing by Deep Reinforcement Learning

Vehicular Crowdsensing (VCS) is a pivotal component in advancing Intelligent Transportation Systems (ITS), facilitating the collection and synthesis of extensive data from distributed vehicular networks. Despite its potential, optimizing participation and data acquisition in VCS is challenged by the self-interested nature of individual participants. In this paper, we propose a novel incentive mechanism for VCS, specifically designed to integrate social benefits within a Vehicle Social Network (VSN). A Small-World (SW) network is introduced to model the VSN, providing a more realistic representation of vehicle interactions and enhancing information propagation. VSN enriches the data utility by sharing data within these networks and acts as a non-monetary incentive that is determined by the strength of connections among participants within the constructed SW networks, sustaining participant engagement even with relatively low monetary rewards. We model the VCS campaign as a Markov Decision Process (MDP) that enables vehicles to independently determine their optimal sensing and communication strategies under the SW networks clustering coefficient $K$ and the rewiring probability $p$. To maximize individual utility under incomplete information, we introduce a multi-agent Deep Reinforcement Learning (DRL) approach called IM-SJSC that utilizes Variational Autoencoder (VAE) and Proximal Policy Optimization (PPO) for accurate decision-making processes. Simulation results in the T-Drive real-world dataset validate the efficacy of the proposed approach, showing that the average utility outperforms other baseline algorithms by 25.00%, 54.07%, 145.25%, and 181.82% under varying numbers of vehicles. The proposed algorithm also achieves significant performance improvements in other scenarios, such as different numbers of tasks and varying task basic prices.