AI technologies hit the Go World
By Pavol Lisy | News | 05.02.2016 9:25| Views: 5817
AI technologies hit the Go World
Last week, shocking news have hit the Go World. A Go-program called AlphaGo, developed by Google DeepMind beat our 3-times European Champion Fan Hui  5 : 0  in an official match played in October 2015. This is the first time that a computer beat a professional Go player in an even match. Most people expected this to happen 5-10 years later, that is why everyone is so surprised.

    Fan Hui and AlphaGo played all over 10 matches (5 formal, 5 informal), where Fan Hui managed to beat the program in 2 out of 5 informal matches. Overall score 8:2 for AlphaGo, but it was agreed before the match that only the formal games will count for the final score. The thinking time for the formal games was 1 hour Basic Time + 3x30 seconds Byo-yomi. For the informal games it was the same, but without Basic Time.

Game 1: AlphaGo (white) won by 2.5 points
AlphaGo is sometimes "rude", it plays moves that have the highest probability to win the game.

    AlphaGo is going to challenge one of the best players nowadays and the best player of last decade, Lee Sedol 9p from South Korea. The 5-game challenge match will be held in Seoul, South Korea 9th-15th March. Match days: 9th, 10th, 12th, 13th, 15th March. The prize for the winner is $ 1.000.000.

    We asked a few questions from a Computer-Go expert Petr Baudis from Czech Republic, the creator and maintainer of PachiBot, one of the best well known Computer programs, also mentioned in AlphaGo's paper.

EGN: "How do you personally feel about this breakthrough in AI technologies?"

Petr: "Like about everyone in the Computer Go world, I am very surprised! I think noone expected such a breakthrough so soon. 
    Another point of surprise is that while there are fundamental improvements in AlphaGo relative to what was done before, they are not related to what we thought were the fundamental problems - namely, the inability of the MCTS programs to focus on a complicated situation (like a tsumego or semeai), properly solve it and (crucially) remember the solution while the board keeps changing.
    Even so, this represents a huge scientific advancement, and surprises are the most exciting part of science!"

EGN: "How much does the Neural Network help from the overall performance of AlphaGo program? And how important is the Hardware in combination with MCTS in your opinion?"

Petr: "The Neural Networks represent the "intuition" of AlphaGo. As any Go player knows, shape intuition is the very basis of strength, and the AlphaGo neural networks have seen tens of millions of Go games, learning about what kind of moves work or don't.
    Neural Networks are making a big comeback in AI in the last 5-10 years and they are improving many aspects of machine learning, ranging from speech recognition (on your phone or in YouTube videos) to image processing (like Google showing related images on the same topic, or Facebook recognizing your friends on photos).  AlphaGo isn't the first time they were applied - recent successes include University of Edingborough, Google/DeepMind, Facebook as well as the oakfoam open source bot. But AlphaGo now introduced a much smarter way to train a "value network" that, rather than directly recommending the next move to make, judges a board position as good or bad (positional judgement) - and this unlocked great power!  But there is another aspect that is probably equally important, but less emphasized - DeepMind was able to apply huge clusters to train the neural networks (doing the same thing on a home gaming rig could take 3-5 years of nonstop number crunching, I'd estimate), and apparently the increased accuracy of the network resulted in a qualitative shift in the way the program understands Go. There are more discoveries behind the curtain here!
    So, just as a strong player can play 5 simultaneous games with weaker players almost purely by intuition - even without any game tree search (that is, explicit reading of sequences), AlphaGo consistently beats all other bots!  That's the power of Neural Network intuition."

EGN: "If you would be in team with Lee Se Dol , what would you recommend him to let him have the highest chance to win against AlphaGo? I guess, playing moves that are out of the program's database might be useful. Trying not to play joseki, play some weird moves that the computer has never seen in the beginning could be useful? Do you have something else?"

Petr: "This stems from the two aspects I have explained so far - the program has extremely strong (maybe even above pro-level?) shape intuition, but the fundamental reading problems of MCTS bots remain unsolved.
    The difference to old programs without the value network is that it can probably memorize many tsumego shapes - so, find a way to make a position where shape memory does not help, perhaps an uncommon throw-in corner shape, tricky joseki deviation, or just a multi-move semeai (and just leave it looming on the board, driving the bot into despair move by move).
    This is a bit of an extrapolation by me and a personal opinion - perhaps the network will always steer the game in the calmer direction, or has deeper capabilities than apparent - but if Lee Sedol finds himself some Computer Go expert to explain everything, I'll bet on him any time.  And I believe that strong players would soon find systematic weaknesses that allow them to start beating the program after a few "probing games".

EGN: "Do you have any plans with PachiBot? Will it compete with alphago?"

Petr: "PachiBot is currently not very actively developed - I have moved on to other things so to evolve, it needs some new volunteers!  But other Go programmers are not sitting idle.  It was completely overshadowed by AlphaGo, but on the very day of this announcement, the Zen team introduced the strongest program so far to KGS - Zen19X uses neural networks too and holds a very solid 7d rank (first bot achieving that). For Computer Go programmers, the game is still on!"

    We also asked a few questions from Fan Hui, soon an interview with him will appear on the website!

    I personally think that this news is positive for the Go World. As the Google Deepmind team members say, I also share their opinion and many Go players think the same, that this can help to popularize the game of Go more in the western countries and all over the world.
    Another positive thing is that in the future, I guess, we will be able to learn from computer's and move closer and closer to perfect play.
    Also, I don't know if it will happen any soon, but I believe it will happen one time, that the game will be "solved" and we will see which move is the best first choice, what is the real amount of komi and other interesting things that only the computers can say for sure.

We all look forward to the AlphaGo match against Lee Sedol in March!
No comments posted...
Leave a Comment
* Name
* Email (will not be published)
Please, enter the name of go in korean!
* - Required fields