Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Google reveals secret test of AI bot to beat top Go players

nature.com

Post reply on HN
Print view View original

31–40 of 214 posts

Re: Google reveals secret test of AI bot to beat top Go players

#31

Post by rictic »

jkkramer wrote: post #26
dfan wrote: post #7

One thing that isn't made clear in this writeup is that Master plays in a very nonhuman style, as opposed to the version of AlphaGo that beat Lee Sedol, which mostly played like a strong human except for a few surprising moves. My first guess when I saw Master's games was that it was a program like AlphaGo that had its policy network trained from scratch rather than being bootstrapped by being given the goal of imita…

I would call its style unorthodox rather than nonhuman. It still plays common josekis (standard opening sequences) but often chooses uncommon variations. Its mid game is full of startling moves backed by VERY good reading. There's definitely still discernible strategy that us mortals can learn from. If I recall correctly, the version that beat Lee Sedol was trained on amateur games plus self-play. My guess would be t…

The version that beat Lee Sedol was trained on pro games.

Re: Google reveals secret test of AI bot to beat top Go players

#32

Post by kefka »

jaxondu wrote: post #24
koolba wrote: post #8

This seems like the beginnings of the plot of an anime I'd want to watch. Season two would probably start with the developer's getting complacent and a Chinese AI entering the scene.

Really hope to see one machine against another. In future we'd probably have game tournament of AI machines. Example my tensorflow-build machine against your openai-build machine in a StarCraft game.

Corewars, where the program is written by a neural net.

Re: Google reveals secret test of AI bot to beat top Go players

#33

Post by drcode »

bluetwo wrote: post #21
dfan wrote: post #7

One thing that isn't made clear in this writeup is that Master plays in a very nonhuman style, as opposed to the version of AlphaGo that beat Lee Sedol, which mostly played like a strong human except for a few surprising moves. My first guess when I saw Master's games was that it was a program like AlphaGo that had its policy network trained from scratch rather than being bootstrapped by being given the goal of imita…

I don't know much about Go but I thought a comment about AlphaGo was interesting that it played in a way to marginally beat the player, which was different that most masters played, which is to clearly beat the opponent by as wide a margin as possible. Is this accurate? Does MasterP also use this style? Are there humans that can play this way? (I'm asking you because you seem to know what you are talking about here.)

I think this is somewhat incorrect- The creators of AlphaGo made it clear that their system does not take the opponent into account at all, it just answers the question "What is the strongest move right now?" and plays that move, without taking the opponent into account. In other words, it does not have any mental model of the opponent.

However, you are correct insofar that it doesn't care about winning by large margins, it prefers winning by smaller margins if it can achieve this with a higher win probability.

Re: Google reveals secret test of AI bot to beat top Go players

#34

Post by jkkramer »

rictic wrote: post #31
jkkramer wrote: post #26

Earlier quoted context omitted.

I would call its style unorthodox rather than nonhuman. It still plays common josekis (standard opening sequences) but often chooses uncommon variations. Its mid game is full of startling moves backed by VERY good reading. There's definitely still discernible strategy that us mortals can learn from. If I recall correctly, the version that beat Lee Sedol was trained on amateur games plus self-play. My guess would be t…

The version that beat Lee Sedol was trained on pro games.

Got a source? I can only find references to training on amateur games (e.g. https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol#Alpha...)

Re: Google reveals secret test of AI bot to beat top Go players

#35

Post by drcode »

jkkramer wrote: post #26
dfan wrote: post #7

One thing that isn't made clear in this writeup is that Master plays in a very nonhuman style, as opposed to the version of AlphaGo that beat Lee Sedol, which mostly played like a strong human except for a few surprising moves. My first guess when I saw Master's games was that it was a program like AlphaGo that had its policy network trained from scratch rather than being bootstrapped by being given the goal of imita…

I would call its style unorthodox rather than nonhuman. It still plays common josekis (standard opening sequences) but often chooses uncommon variations. Its mid game is full of startling moves backed by VERY good reading. There's definitely still discernible strategy that us mortals can learn from. If I recall correctly, the version that beat Lee Sedol was trained on amateur games plus self-play. My guess would be t…

> My guess would be that this new version relies more heavily on pro games.

Unlikely, since AlphaGo can now generate large numbers of "pro quality" games from scratch. I think it's far more likely it is an autodidact at this point.

Re: Google reveals secret test of AI bot to beat top Go players

#36

Post by Ajedi32 »

jaxondu wrote: post #24
koolba wrote: post #8

This seems like the beginnings of the plot of an anime I'd want to watch. Season two would probably start with the developer's getting complacent and a Chinese AI entering the scene.

Really hope to see one machine against another. In future we'd probably have game tournament of AI machines. Example my tensorflow-build machine against your openai-build machine in a StarCraft game.

There's a pretty interesting Starcraft AI tournament going on right now actually, though I'm not certain whether any of the bots are using deep learning techniques: http://sscaitournament.com/

Re: Google reveals secret test of AI bot to beat top Go players

#37

Post by natosaichek »

igravious wrote: post #19
natosaichek wrote: post #17

I'm pretty sure this[1] is the archives of the bot's play. You can download the SGF files of the matches from there and view them online at EidoGo[2] [1] http://www.gokgs.com/gameArchives.jsp?user=Master [2] http://eidogo.com/upload

Marcel Grünauer over on LifeIn19x19 has made a compilation of all 60 of Master's games. Found it just then! (A zip archive of 60 sgf[1] files) I _think_ that's all of them as of now. http://lifein19x19.com/forum/viewtopic.php?f=18&t=13935 [1] https://en.wikipedia.org/wiki/Smart_Game_Format ps: The Japanese term for a game record for the game of Go is kifu: https://en.wikipedia.org/wiki/Kifu Hence http://gokifu.com/ w…

Ah! Thanks for your evidence in support of cunningham's law[1]. I'm delighted to have access to the real deal.

[1] https://meta.wikimedia.org/wiki/Cunningham%27s_Law

Re: Google reveals secret test of AI bot to beat top Go players

#38

Post by hyperpape »

gertef wrote: post #28
bluetwo wrote: post #21

Earlier quoted context omitted.

I don't know much about Go but I thought a comment about AlphaGo was interesting that it played in a way to marginally beat the player, which was different that most masters played, which is to clearly beat the opponent by as wide a margin as possible. Is this accurate? Does MasterP also use this style? Are there humans that can play this way? (I'm asking you because you seem to know what you are talking about here.)

Humans play for a large lead because they don't have enough memory/power to accurately estimate the value of their positions, so they play for a buffer -- AlphaGo has higher confidence in its valuation, so it can play it closer -- ~85% confidence of winning by 5 stones (with room for error) vs 99% chance of winning by 2 stones

I believe that accuracy/"self-confidence" is part of it. However, I think it's also the case that AlphaGo has a monte carlo tree search in addition to the neural net, so it sometimes plays more conservatively than it needs to because it overweights obscure possibilities ("defending here is not necessary, but by doing so, I prevent some number of playouts where I play a dumb move and lose, and I can still win even if I defend").

Humans do the same thing, playing conservatively in a situation where they're far enough ahead. The difference is that a human sometimes looks at a move and says "This move works, and gains points. There is no risk." For a bot using MCTS, everything is a probability.

Re: Google reveals secret test of AI bot to beat top Go players

#39

Post by hyperpape »

jkkramer wrote: post #34
rictic wrote: post #31

Earlier quoted context omitted.

The version that beat Lee Sedol was trained on pro games.

Got a source? I can only find references to training on amateur games (e.g. https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol#Alpha... )

No, but I'm pretty sure the parent post is right. Pro games were included.

Edit: I'm not actually that sure. I'm asking around right now (with go players, not DeepMind people).

Re: Google reveals secret test of AI bot to beat top Go players

#40

Post by hyperpape »

badminton1 wrote: post #29
hwc wrote: post #6

Earlier quoted context omitted.

Elo Ratings are unbounded.

Yes but players are usually presented by their rank, e.g: Ke Jie 9P, Lee Sedol 9P. Then, professional dans in some Go associations are granted by total wins over their career rather than relative strength to another rank (e.g: Japan).

There is already a wide range of strength among 9p players. In addition, once a player reaches 9p, they never lose it, so it includes players in their prime as well as players who have declined.
Post reply on HN