Resumen de Mobile crowdsourcing with task transfers: a market-based multi-agent approach
Jeremias Dötterl
In mobile crowdsourcing, spatial tasks are outsourced to an on-demand workforce of autonomous individuals who find nearby tasks using their smartphones. The crowd workers, spread across the city, can perform many tasks quickly and efficiently, like collecting sensor data from different locations or delivering parcels to nearby destinations. However, operating mobile crowdsourcing systems reliably is difficult.
Tasks get frequently assigned to crowd workers who during task execution struggle to complete these tasks successfully, causing high failure rates and low service quality. A promising remedy, which has been exploited successfully in related domains, is to continuously adapt the assignment and to respond to failure-provoking events by transferring tasks to better-suited workers who use other routes or vehicles. However, introducing task transfers to mobile crowdsourcing is difficult because workers are autonomous and may reject transfers that, though globally efficient, are not beneficial to them individually. Furthermore, task outcomes are uncertain and must be predicted. Also, in crowdsourced delivery, transfers require the handover of parcels, which makes the coordination of task transfers even more challenging.
This dissertation studies the problem of coordinating task transfers among autonomous self-interested crowd workers. The dissertation presents a formal model of that problem and proposes a market-based solution that allows crowd workers, represented by software agents, to trade tasks on a computational marketplace. Using auctions, the agents sell and buy tasks according to the workers' situations and preferences while simultaneously reducing the risk of task failures. A new architecture is presented that allows agents to make situation-aware offer and bidding decisions. The agents apply machine learning on sensor data obtained from the workers' mobile devices to predict task outcomes and to exploit the predictions in their decisions.
Simulation experiments show that the approach can reduce task failures effectively and help to operate mobile crowdsourcing systems more reliably.
