Multi-objective Optimization Methods for Allocation and Prediction


Speaker


Abstract

In this thesis we focus on two different aspects of auctions and we employ techniques and methods from both operations research and computer science.

First, we study the allocation of tasks to agents at the end of an auction. Usually, tasks are allocated in such a way that minimizes the total cost for the auctioneer. This allocation is optimal in a one-shot auction, but if the auction is repeated, this can have negative consequences for the results in the long run. Therefore, we consider a fair allocation, which costs slightly more in a one-shot auction, but has positive effects on the participation level of agents and the total cost for the auctioneer in repeated auctions.

Second, we consider the auction design. How an auction is set up, like which tasks should be auctioned first, or what the starting price should be, impacts the result. Usually there are experts who know what has occurred in previous auctions and how a future auction should be designed in order to obtain the best results. However, historical auctions can obtain so much information that experts overlook things. We use a combination of machine learning and optimization models to extract information from historical auctions and use this information to help design future auctions for better results.