Adaptive task recommendation based on reinforcement learning in mobile crowd sensing

Adaptive task recommendation in Mobile crowd sensing (MCS) is a challenging problem, mainly because perceptual tasks are spatio-temporal in nature and worker preferences are dynamically changing. Although there have been some approaches to address the dynamics of task recommendation, these approaches suffer from several problems. First, they only learn the worker’s past preferences and cannot cope with the situation where the worker’s preferences may change in the next moment, and they only consider the current optimal recommendation instead of global optimization. Second, existing methods do not scale efficiently to the arrival of new workers or tasks, requiring the entire model to be retrained. To address these issues, we propose an adaptive task recommendation method (ATRec) based on reinforcement learning. Specifically, we formalize the adaptive task recommendation problem for each target worker as an interactive Markov decision process (MDP). Then, we use an improved matrix decomposition technique to construct worker-personalized latent factor states based on information such as task content and spatio-temporal context, enabling us to use a unified MDP to learn optimal strategies for different workers. After that, we design an adaptive update algorithm (AUA) based on Deep Q Network (DQN) to more accurately learn the dynamic changes of workers’ preferences to adaptively update the task recommendation list of workers. In addition, we propose a personalized dimension reduction method (PDR) to reduce the size of the task set. Through comprehensive experimental results and analysis, we demonstrate the effectiveness of the ATRec approach. Compared with existing methods, ATRec can better solve the problem of adaptive task recommendation, and can more accurately predict workers’ preferences and make recommendations.