Cost-Aware Bandits Optimize Crowdsensing Worker Recruitment

Mobile crowdsensing (MC) platforms rely on recruiting mobile users to perform sensing tasks, but a key challenge is efficiently managing worker recruitment under conditions of uncertainty. Existing methods often assume a static worker performance, which doesn't reflect the reality that worker quality typically improves with experience before stabilizing. Furthermore, the actual cost incurred by workers can be unpredictable due to varying device and context states. This research tackles these complexities by proposing a more dynamic and cost-aware approach. The problem is framed as a structured bandit model, where the platform must select one worker per round, observe their sensing quality and cost, and repeat until a budget is exhausted. Crucially, each worker's expected sensing quality is modeled as an increasing-then-converging function of their participation count, reflecting an "experience curve." Simultaneously, each worker has an unknown expected cost. The developed cost-aware online learning framework is designed to jointly learn these evolving reward trajectories and heterogeneous costs, detect when performance saturates, and strategically allocate the limited budget to maximize the overall long-term sensing utility. The paper provides theoretical performance guarantees for its approach and validates it through extensive experiments. These experiments demonstrate consistent improvements over baseline methods that either neglect the experience-driven dynamics of worker performance or assume prior knowledge of costs. This framework offers a more realistic and effective solution for managing crowdsensing operations.