Dual-Stage DRL-Based Trust-Aware Task Bundling and Allocating for Mobile Crowdsourcing

Ensuring high data quality in Mobile Crowdsourcing (MCS) is particularly challenging for unpopular tasks, which often suffer from low completion rates under limited budgets. Simply increasing rewards can improve participation but leads to excessive platform costs, highlighting the need for intelligent and cost-efficient task organization mechanisms. To address this challenge, we propose a Dual-stage Reinforcement Learning based Trust-Aware Task Bundling and Allocation (DRL-TTBA) strategy that jointly optimizes task completion rate, data quality, and platform profit. Specifically, task bundling is optimized through a dual-stage deep reinforcement learning framework based on Deep Q-Networks (DQN) and Advantage Actor-Critic (A2C) algorithms, while task allocation adopts a trust-aware strategy that jointly considers worker willingness and trust. A Bayesian trust model with Gaussian-kernel deviation weighting is further introduced to evaluate unknown workers and progressively identify trusted and high-quality workers, enabling a closed-loop interaction between task execution outcomes and subsequent bundling and allocation decisions. Extensive simulations demonstrate that the DRL-TTBA consistently outperforms existing approaches, achieving data quality improvements of 14%~106% while maintaining high task completion rates and improving platform profitability. These results confirm the effectiveness of integrating trust-aware with task bundling for reliable and efficient mobile crowdsourcing.