“Mastering the Game of Go With Deep Neural Networks and Tree Search”, 2016-01-28 (; similar):
The game of Go has long been viewed as the most challenging of classic games for artificial intelligence owing to its enormous search space and the difficulty of evaluating board positions and moves. Here we introduce a new approach to computer Go that uses ‘value networks’ to evaluate board positions and ‘policy networks’ to select moves. These deep neural networks are trained by a novel combination of supervised learning from human expert games, and reinforcement learning from games of self-play. Without any lookahead search, the neural networks play Go at the level of state-of-the-art Monte Carlo tree search programs that simulate thousands of random games of self-play. We also introduce a new search algorithm that combines Monte Carlo simulation with value and policy networks. Using this search algorithm, our program AlphaGo achieved a 99.8% winning rate against other Go programs, and defeated the human European Go champion by 5 games to 0. This is the first time that a computer program has defeated a human professional player in the full-sized game of Go, a feat previously thought to be at least a decade away.
[Anecdote: I hear from Groq that the original AlphaGo GPU implementation was not on track to defeat Lee Sedol by about a month before, when they happened to gamble on implementing TPUv1 support. The additional compute led to drastic performance gains, and the TPU model could beat the GPU model in ~98⁄100 games, and the final model solidly defeated Lee Sedol. (Since TPUv1s reportedly only did inferencing/forward-mode, presumably they were not used for the initial imitation learning, or the policy gradients self-play, but for generating the ~30 million self-play games which the value network was trained on (doing regression/prediction of ‘board → P(win)’), requiring no state or activations from the self-play games, just generating an extremely large corpus which could be easily used by GPU training.)]