|Title||:||Disovering Economic Mechanisms through Machine Learning|
|Speaker||:||Harikrishna Narasimhan (Harvard University)|
|Details||:||Fri, 21 Apr, 2017 12:00 PM @ BSB 361|
|Abstract:||:||The last decade has seen enormous synergy between computer science and economics, providing practical solutions for a number of commercial and social applications, whether it is large-scale internet auctions to sell products on eBay, auctions to sell wireless spectrum to telecom companies, matching medical students to hospitals for internships, or pairing donors for kidney exchanges. One of the cornerstone problems in this area is the design of revenue-optimal auctions. Despite the remarkable progress made over the years, only specialized versions of this problem are well-understood, and the design of optimal auction for even the simple setting of two buyers and two items remains open. On the other hand, the last decade has also seen tremendous success in the use of data-driven machine learning approaches for a variety of prediction tasks, achieving near-human accuracy in applications like speech recognition and machine translation. Can a similar use of data and computation allow us to design practical auctions for problems beyond the reach of theoretical analysis?
In this talk, I will describe our recent work on the use of deep machine learning methods for designing revenue-optimal auctions from data. We model an auction as a multi-layer neural network, and optimize its parameters using a novel loss function that promotes high revenue and low incentive violations. We shall see that neural networks are able to recover known auction designs as well as discover new designs for poorly-understood settings, thus demonstrating the potential that machine learning approaches have in advancing the state-of-the-art. I will also discuss machine learning techniques for designing more general economic mechanisms, and conclude by highlighting broader connections between machine learning and social science.
Speaker Bio: Harikrishna Narasimhan is a post-doctoral fellow at the Institute for Applied Computational Science at the School of Engineering and Applied Sciences, Harvard University. Prior to this, he received his PhD from the Indian Institute of Science, Bangalore, supported by the Google India PhD fellowship. His research interests span the areas of machine learning, statistical learning theory, and optimization, and their connections to economics and social science.