ECE Seminar: Learning and Decision Making in Strategic Environments

Sequential decision problems are everywhere: managing an investment portfolio, routing packets through a network, distributing resources within an organization. The statistical approach to decision-making under uncertainty is to model the world as stochastic and oblivious and to act under IID assumptions. However, in competitive environmentsthese assumptions break down. We will consider a framework in which we are sequentially optimizing a function with non-stochastic or even adversarial inputs and we ask: can we hope to achieve any reasonable guarantees? The resulting story, we will show, is quite positive.The talk will explore several tools from the competitive online learning literature, and we will also dive into a few of nice applications:* Pricing of prediction market contracts + crowdsourcing* Adaptive routing with limited information* Optimization in the "bandit setting"* Hedging options and other financial derivatives, and the relationship to the "Black-Scholes formula"http://www.seas.upenn.edu/~jaber/