Title: Towards Machine Learning Enhanced Combinatorial Auctions?
Abstract: Combinatorial auctions are at the core of Economics and Computation research since, in these mechanisms, efficient representations and processes of the agents' reported value functions are fundamental to achieve desirable performances in large domains. Indeed, certain auction designs may require each agent to reveal an amount of information that grows exponentially in the number of the auctioned items, thus leading to infeasible winner determination problems.
In the last decades, particular iterative combinatorial auctions have been developed to decentralize the computational burden of the mechanism to the agents: the idea is to make them report only the information relevant to determine the efficient allocation. This information is obtained through particular queries, where each agent observes a vector of prices established by the mechanism and reports her favorites bundles at these prices. However, in this scenario, the computational effort required from the agents may become too large since, at each round of the auction, each of them has to process all her value function to correctly reply to such a query. This problem was highlighted in the experimental tests performed by Scheffel, Ziegler, and Bichler in 2012. Here it was observed that the main reason of the outcome inefficiency of some iterative combinatorial auctions was the limited computational capability of the agents, while their strategic behavior and the particular auction design had minor effects. In this talk I will introduce some ideas on how machine learning techniques can address these computational issues.