Designing Incentives for Crowdsourced Tasks via Multi-armed Bandits
Itoh Akihiko, Shigeo Matsubara · 2016
We have examined to apply the multi-armed bandit (MAB) techniques to the incentive design problem in crowdsourcing. So far, researchers studied the effects of various incentives including financial and social ones. The effect of these incentives, however, may differ in different tasks and/or at a different time, which makes difficult to find an appropriate incentive out of the whole design space of incentives is difficult. To overcome this difficulty, we take a data-driven approach instead of trying to find a universally effective incentive. More specifically, we apply MAB algorithms to the incentive design problem. MAB problems are a theoretical model of exploration-exploitation tradeoffs in reinforcement learning. We evaluated the effects of the nine incentives by using the task of word search puzzles published on Amazon Mechanical Turk. Based on the obtained data, we further evaluated the performance of typical MAB algorithms by simulation. The results show that applying the MAB algorithms are effective to obtain the appropriate incentive design.