Learning when to Auction and when to Bid

The market based approach is widely used to solve the problem of multirobot coordination. In this approach, communication and computation costs are key issues, but have not been carefully addressed by the different architectures in the literature. In this paper, we present a method to reduce these costs, by adding the capability to learn whether a task is worth offering up for auction and also whether it is worth bidding for the task, based on previous experience about successful and unsuccessful bids. We show that the method significantly decreases communication and computation costs, while maintaining good overall performance of the team.