Allocation of complex crowdsourcing tasks, which typically include heterogeneous attributes such as value, difficulty, skill required, effort required and deadline, is still a challenging open problem. In recent years, agent-based crowdsourcing approaches focusing on recommendations or incentives have emerged to dynamically match workers with diverse characteristics to tasks to achieve high collective productivity. However, existing approaches are mostly designed based on expert knowledge grounded in well-established theoretical frameworks. They often fail to leverage on user-generated data to capture the complex interaction of crowdsourcing participants’ behaviours. This paper aims to address this challenge.

The paper proposes a policy network plus reputation network (PNRN) approach which combines supervised learning and reinforcement learning to imitate human task allocation strategies which beat artificial intelligence strategies in this large-scale empirical study. The proposed approach incorporates a policy network for the selection of task allocation strategies and a reputation network for calculating the trends of worker reputation fluctuations. Then, by iteratively applying the policy network and reputation network, a multi-round allocation strategy is proposed.

PNRN has been trained and evaluated using a large-scale real human task allocation strategy data set derived from the Agile Manager game with close to 500,000 decision records from 1,144 players in over 9,000 game sessions. Extensive experiments demonstrate the validity and efficiency of computational complex crowdsourcing task allocation strategy learned from human participants.

The paper can give a better task allocation strategy in the crowdsourcing systems.