Computer Science Seminar: On Learning in Markets: Dealers and Dating
Sanmay Das
MIT
Department of Electrical Engineering & Computer Science
Faculty Host: Victor Lesser
"On Learning in Markets: Dealers and Dating"
Artificial agents placed in market environments face difficult learning and decision-making problems. "Rational" behavior is hard to specify, and the design of good algorithms for agents that operate in these environments presents an important challenge for computer science. In addition to their practical value, such algorithms could also be leveraged to better understand market outcomes. Traditionally, the fields of economics and finance theory simplify their models until agent decision problems are easily solved. Algorithms that work in more complex situations would allow us to build richer models and study the dynamics of economic and social systems. I will illustrate this research agenda of designing successful algorithms for complex market settings and subsequently using these algorithms to study market properties with two specific examples from my research.
First, I will present an algorithm for setting prices in a dealer-driven stock market (like the New York Stock Exchange). The algorithm uses a Bayesian non-parametric density estimation technique to maintain an estimate of the underlying fundamental value of a stock and approximately solve the price-setting equations. I will show how we can use this algorithm to study price properties in the model in a manner that was previously impossible.
Second, I will discuss the decision problems faced by agents in repeated matching environments with learning in the context of a "dating market" in which men and women repeatedly go out on dates and learn about each other. The focus of this part of the talk will be on the effects of three different mechanisms for pairing agents, especially in terms of the asymptotic stability of the resulting matchings when agents use a simple learning rule coupled with an epsilon-greedy exploration policy.
I will conclude by examining the important problems for this area of research and highlighting the connections with different parts of computer science, economics, and the brain and cognitive sciences.
(Part of this talk is based on joint work with Emir Kamenica.)
