How Facebook’s AI Researchers Built a Game-Changing Go Engine

ColinWright · on Dec 7, 2015

Comment[0] by fitzwatermellow[1] yesterday from one of the other submissions[2]:

    Here's the paper:

    Better Computer Go Player with Neural Network and
    Long-term Prediction by Yuandong Tian, Yan Zhu

    http://arxiv.org/abs/1511.06410[3]

    "Adding MCTS to darkforest creates a much stronger
     player: with only 1000 rollouts, darkforest+MCTS
     beats pure darkforest 90% of the time; with 5000
     rollouts, our best model plus MCTS beats Pachi
     with 10,000 rollouts 95.5% of the time."

[0] https://news.ycombinator.com/item?id=10685334

[1] https://news.ycombinator.com/user?id=fitzwatermellow

[2] https://news.ycombinator.com/item?id=10685278

[3] http://arxiv.org/abs/1511.06410

kawera · on Dec 7, 2015

Thanks, I didn't see it yesterday otherwise wouldn't have posted.

pmontra · on Dec 7, 2015

The approach is novel but at the moment it's not stronger than the best traditional go programs. See http://senseis.xmp.net/?ZenGoProgram for possibly the very best one.