Linde Institute/SISL Seminar: Kevin Leyton Brown
Modeling Human Play in Unrepeated Games
This talk will survey our ongoing work on modeling human play of unrepeated, simultaneous-move games, both communicating some of the key lessons we've learned and giving previews of the new directions we're currently pursuing.
I'll begin by surveying our published work, in which we've performed a large-scale meta-study of previous work from Economics, evaluated existing models from the literature, and ultimately proposed a variation on one of these models that achieves good across-the-board performance. We've also proposed a new framework for describing nonstrategic agents which improves the performance of our models overall.
I'll also describe the ideas behind two avenues of ongoing research. First, existing behavioral models ignore the fact that people think more strategically in some situations than others; we've found a way to capture this, and have some initial evidence that we're on the right track. Second, deep learning is a machine learning paradigm with many recent successes; I'll explain why it's not straightforward to apply to our setting, why it's desirable to do so anyway, and how we can achieve state-of-the-art performance by doing so.
Bio and photos: http://www.cs.ubc.ca/~kevinlb/bio.html.
Contact: Barbara Estrada at 626-395-4083 firstname.lastname@example.org