Computers reach 5d on KGS

[yoyoma]

Computer Go continues to improve, and recently two programs on KGS have reached 5d on KGS! First CrazyStone, using the account 'bonobot', although it disappeared after a few days and drifted back down to 4d. Next Zen19D appeared and reached 5d, and it seems to be holding that level consistently.

How long do you think it will be until the first bot reaches a consistent 6d on KGS? It was only August 2010 that Zen was the first bot to consistently hold a 3d rating, and now less than a year later it is 5d!

Here is a history of bot ratings over time: http://senseis.xmp.net/?KGSBotRatings

[jts]

I'm confused. If so, why couldn't they get a computer to beat that low-dan player this past winter? Are the KGS bots using extremely expensive hardware?

[daniel_the_smith]

Are they playing with low time limits?

[hyperpape]

Part of the idea behind the challenge was that a human could learn the computer's weaknesses, thus beating a bot that might maintain a higher rating against a large field of (undedicated) opponents.

Also, these bots are not the one that was chosen, though I don't know if they're substantially better than MFoG or not.

An older thread had a great deal of discussion: viewtopic.php?f=18&t=2646&hilit=shodan+bet

[yoyoma]

Yes these bots are playing with fast times -- 9x:15. Typically they lose about 1 stone when they play something like 30min main + 5x:30 overtime.

About the Tromp bet, the hardware Zen19D is running on is too expensive to qualify for that bet. Also the time controls would favor Tromp even more than 30min main. And finally a huge factor is Tromp would be playing much more seriously than the typical opponent on KGS!

Despite the caveat about blitz time controls, 5d is still quite amazing! Less than two years ago Zen was 3d under similar time controls.

[EdLee]

Are they playing with low time limits?

Yes.

[hyperpape]

This makes me wonder: are bots allowed on DGS?

[palapiku]

Zen reached 1d only about two years ago. The progress the bots are making is stunning. I believe better-than-human play is not that far off, maybe 10 years.

6d this year won't surprise me.

[ez4u]

I'm confused. If so, why couldn't they get a computer to beat that low-dan player this past winter? Are the KGS bots using extremely expensive hardware?

This is from the Zen19D info... ("me" = someone named Hideki Katoh)
Q: What is Zen19D?
A: Zen19D is a version of Zen19, runs on a mini-cluster of 6 pcs (a 6-core Xeon W5680/4 GHz, two 4-core i7 920/3.2 GHz, and three 4-core Core2Quad/3 GHz) connected via a GbE LAN. Although Zen19 is a prototype of a commercial product, Zen19D is being developed for academic research by team DeepZen, a joint project of ZenAuthor and me.

Q: What does D stand for?
A: Distributed version, or DeepZen.

Q: How strong is it?
A: Estimated 5d for blitz and 4d for longer games

[snorri]

It would be fun to see a Malkovich game with Zen19D if it has any kind of tracing on how it is weighting its moves.

[BobC]

this is IMO very sad.... Chess lost all sparkle for me once I realised computers were becoming dominant.

[RobertJasiek]

Nothing changes ALA programs do only use sheer calculation power. Humans will have to "justify" their superiority only when computers start to explain their decisions by human-readable reasoning and can maintain their hard- and software alone.

[Mnemonic]

Another prediction I made in The Age of Intelligent Machines was that once computers did perform as well or better as humans in chess, we would either think more of computer intelligence, less of human intelligence, or less of chess, and that if history is a guide, the last of these would be the likely outcome. Indeed, that is precisely what happened. Soon after Deep Blue's victory we began to hear a lot about how chess is really just a simple game of calculating combinations and that the computer victory just demonstrated that it was a better calculator.

[hyperpape]

What goes unmentioned in Kurzweil's comment is that people once thought that once we had a chess playing computer, we would have computers that recognized speech, had vision and could write a novel. But it turned out that chess was vastly easier than those, and really didn't lead to much progress on those fronts. So the earlier valuation of a chess playing computer was based on massive factual errors.

[Mike Novack]

Robert, that's a somewhat naive way to think about intelligence, whether natural or artificial. You are making some assumptions both about the current state of the art and the possibilities.

a) explain "why" in human understandable terms
You are assuming that none of the existing programs can do this? Just because few of the programs have been given this capability doesn't make that so. What I believe isn't possible at the present time is giving a why for "why is move A (which has x, y, z "go reasons" behind it) better than move B (which has u, v, w go reasons behind it)". In other words, in human understandable terms, why in this instance are x, y, and z more important than u, v, and w. This of course assumes that you don't consider "human understandable" the concept that if out of 100 games played between opponents of equal skill from position A if one side wins more often than out of 100 games played from position B then for that side position A is likely better than position B (and even more likely if that is true for 1000 games).

Regardless of how a go playing program has selected X as its next move (the move it thinks best) it could then turn around and "waste time" analyzing X in terms of the possible "go reasons" for the move. Feature could be added even to the "pure" MCTS programs if somebody wanted to so.

b) Some forms of AI are essentially "programming themselves" and cannot possibly explain why/how. To understand this you need to understand that a "program" can be the combination of "code" (process and/or function evaluation) and "data" and the behavior of the program might depend very heavily on the data.

You are perhaps going to assume that this data is supplied by us humans? Not necessarily so. Look up the form of AI that goes by the designation "neural net". Here all that is necessary is some way to judge/evaluate how well the "program" has performed its assigned task (the task it is supposed to learn) and then random peturbations of the data can lead to evolution of better and better performance as the neural net learns to perform the task. Not meaningful in this case to ask "why" this set of data works (other sets might work equally well).

You think not? You think a justification of "why does this set of data work" necessary to prove intelligence? Then kindly explain what about the details of the connections between the nerve cells in your brain lets you perform some task. More than "as I learned the task connections were broken and made until now, with what remains, it just works".