04-02

Learning in an Adversarial World, with Connections to Pricing, Hedging, and Routing

Machine Learning is often viewed through the lens of statistics, where one tries to model or fit a set of data under stochastic conditions; for example, it is typical to assume one's observations were sampled IID. However, dating back to results of Blackwell and Hannan from the 1950s we know how to construct learning and decision strategies that possess robust guarantees even under adversarial conditions. Within this setting the goal of the learner is to "minimize regret" against any sequence of inputs. In this talk we lay out the framework, discuss some recent results, and we finish by exploring a couple of surprising applications and connections, including: (a) market making in combinatorial prediction markets, (b) routing with limited feedback, and (c) hedging derivative securities (e.g. European option contracts) in the worst-case, with a connection to the classical Black-Scholes option-pricing model.

Jake received his undergraduate degree in Mathematics from MIT in 2002 and a Master's degree in Computer Science from TTI-C in 2006. He recently finished a PhD in Computer Science at UC Berkeley, advised by Peter Bartlett, and he is now the Simons Postdoctoral Fellow at University of Pennsylvania. Jake has a particular focus on the intersection between machine learning, games and markets.