CrowdControl : An online learning approach for optimal task scheduling in a dynamic crowd platform

The dynamic nature of crowd platforms poses an interesting problem for users who wish to schedule a large set of tasks on a given platform. Although crowd platforms vary in their performance characteristics, certain temporal patterns can be discerned and statistically modeled. Methods that can learn these patterns and adapt as the patterns change can schedule “the right number of tasks with the right price at the right time” which can have significant implication on how well the tasks are completed. To address the problem, we propose CrowdControl : a novel online approach that controls and co– ordinates the execution of crowd tasks, in real–time, involving simultaneous learning of crowd performance and optimization based on the learning. We design and compare several algorithms in this framework. We also describe dynamic statistical models of crowd performance based on real data and a simulation testbed for evaluating CrowdControl algorithms. Our experiments show that algorithms that schedule jobs by adaptively learning current crowd performance can significantly outperform other algorithms that do not learn or rely on past data alone. ICML Workshop: Machine Learning Meets Crowdsourcing, Atlanta, Georgia, USA, 2013.