AlphaGo: Mastering the game of Go with deep neural networks and tree search

19^th April 2016, 13:00 Leggate Theatre, Victoria Gallery & Museum
Dr. Marc Lanctot
Google Deepmind

Abstract

The game of Go has been long been seen as a grand challenge problem in artificial intelligence due to the inherent complexity arising from its large branching factor. In this talk, I will describe AlphaGo, which uses a new approach that combines deep learning from expert play and self-play reinforcement learning of a 'value network' to evaluate positions, and a 'policy network' to recommend moves. These networks are then used to inform the simulations of a distributed Monte Carlo tree search. AlphaGo defeats other Go programs 99.8% of the time, and won 5 games to 0 against the European Go champion, marking the first time a computer program has defeated a human expert on the full-sized board. Finally, I will discuss the match in Seoul against the 9-dan professional, Lee Sedol.