Deep Reinforcement Learning Based Iterative Participant Selection Method for Industrial IoT Big Data Mobile Crowdsourcing

With the massive deployment of mobile devices, crowdsourcing has become a new service paradigm in which a task requester can proactively recruit a batch of participants with a mobile IoT device from our system for quick and accurate results. In a mobile industrial crowdsourcing platform, a large amount of data is collected, extracted information, and distributed to requesters. In an entire task process, the system receives a task, allocates some suitable participants to complete it, and collects feedback from the requesters. We present a participant selection method, which adopts an end-to-end deep neural network to iteratively update the participant selection policy. The neural network consists of three main parts: (1) task and participant ability prediction part which adopts a bag of words method to extract the semantic information of a query, (2) feature transformation part which adopts a series of linear and nonlinear transformations and (3) evaluation part which uses requesters' feedback to update the network. In addition, the policy gradient method which is proved effective in the deep reinforcement learning field is adopted to update our participant selection method with the help of requesters' feedback. Finally, we conduct an extensive performance evaluation based on the combination of real traces and a real question and answer dataset and numerical results demonstrate that our method can achieve superior performance and improve more than 150% performance gain over a baseline method.