Information Session | September 17 | 5:30-7 p.m. | Soda Hall, Wozniak Lounge (430)
Accelerating Self-Play Learning in Go
Presented by: David Wu
Over the last few years at Jane Street, we've been increasingly exploring machine learning to build more sophisticated models in various areas of trading, and many of us are fascinated by the developments in the state of the art as a whole over that time.
Around early 2018, I started a personal hobby project to explore neural net training ideas in Go. With Jane Street's help, we've now been able to test at scale and we can improve the learning efficiency of DeepMind's AlphaZero process in Go by a factor of about 50! Many of the improvements reflect principles that we think can apply much more broadly to other problems we care about. In this talk we will discuss how AlphaZero works and the lessons learned in trying to do even better.
Food will be provided!