Learning Complex Crowdsourcing Task Allocation Strategies from Humans

Efficient allocation of complex tasks, which typically include heterogeneous attributes such as value, difficulty, skill required, effort required and deadline, is a challenging open problem in crowdsourcing. Existing approaches are mostly designed based on expert knowledge and fail to leverage on user generated data to capture the complex interaction of crowdsourcing participants' behaviours. In this paper, we propose a data-driven learning approach to address this challenge. The proposed approach combines supervised learning and reinforcement learning to enable agents to imitate human task allocation strategies which have shown good performance. The policy network component selects task allocation strategies and the reputation network component calculates the trends of worker reputation fluctuations. The two networks have been trained and evaluated using a large-scale real human task allocation strategy dataset derived from the Agile Manager game. Extensive experiments based on this dataset demonstrate the validity and efficiency of our approach.