Reproducing AlphaZero: what we learn: BLISS Seminar

Seminar | March 4 | 3-4 p.m. | 540 Cory Hall

 Yuandong Tian, Facebook AI Research

 Electrical Engineering and Computer Sciences (EECS)

We reproduce and open source AlphaGoZero/AlphaZero framework using 2000 GPUs and 9 days, achieving super-human performance of Go AI that beats 4 top-30 professional players with 20-0, provide extensive ablation studies and perform basic analysis. In this talk we will share our journey and interesting first-hand experience that makes a large-scale RL system work. We hope it will spur future research both practically and theoretically.