Lunch & Learn
Duke University Computer Science Researcher Michael Albert will be visiting UVA to give a lunch & learn talk titled, "Data-Driven Mechanism Design."
Friday, March 16 from Noon to 1:30pm
Location: Dell 1 Building, Room 105
Many natural settings of interest involves groups of self-interested agents competing for resources. The design and optimization of these systems is the field of mechanism design.
However, there are well-known impossibility results in mechanism design related both to revenue maximization in thin markets and revenue minimization, while maintaining social efficiency, in general settings. These impossibility results have significant negative implications for many settings, including online advertising auctions and federated server farms.
Albert will discuss how we can combine large scale data about the participants (or agents) of the mechanism, potentially gathered from previous rounds in a repeated setting, to sidestep these impossibility results. We do this by learning a tailored mechanism from the data using techniques from machine learning, robust optimization, and economics. This work provides a computationally efficient and sample efficient algorithm that learns from historical data and, in simulation, leads to significantly (both statistically and economically) increased revenue.
As an additional application, Albert will demonstrate that these techniques can lead to mechanisms that allocate server resources that are nearly zero expected payment while maintaining social efficiency using a panel data set of server utilization for jobs on a Google cluster, significantly outperforming existing techniques. This result opens the door to fully distributed resource allocation systems.
Albert will also discuss applications of data driven mechanism design to optimize vehicle routing in a congested traffic network and efficient scheduling of charge distribution in a congested electric grid with applications to energy pricing.
For more information, contact Savanna Galambos.