Did GNUGo get much better?

[ramanujan]

Has there been an improvement in its programming? Last year when I would play it even, I would win by 50 points. Now it's beating me %50 of the time.

[topazg]

Yes, it seems to have gotten much much better. I used to play it at 6 stones and find the game relatively easy, now I have a slightly under 50% win-rate with 6 stones, not that I play it a great deal.

[Mike Novack]

I'm not sure what you are saying. How are you playing against gnugo?

a) This is "free software" and readily available. So you could, be playing against it on your own computer. In which case, unless you changed whatever version was installed you shouldn't experience a change in how it plays. There are sorts of AI that can "learn" but this wouldn't be a case of that.

b) If you are playing against a bot on some server this could be a change in what version of gnugo was installed on that computer. Not necessarily a change in what versions of gnucash are available but a change in what version actually installed. Some people like to always install the newest and "best" while others tend to wait till all warts removed from a new version (aka "avoid the bleeding edge).

c) Theoretically when playing against a computer program you should win by a large amount or lose by a small amount. If a program predicts that playing normally it will lose by a small amount its only rational choices are "resign" or "take any chance to pull the game out". A human player might take solace from a loss "at least it was close" but we humans aren't entirely rational (computers have no emotions at all). So going by margin of victory can be very misleading. Don't assume "if I am losing by only ten points I will win if I take one more handicap stone". Not unlike when a very strong player is teaching us.

[ethanb]

I'm not sure what you are saying. How are you playing against gnugo?

a) This is "free software" and readily available. So you could, be playing against it on your own computer. In which case, unless you changed whatever version was installed you shouldn't experience a change in how it plays. There are sorts of AI that can "learn" but this wouldn't be a case of that.

b) If you are playing against a bot on some server this could be a change in what version of gnugo was installed on that computer. Not necessarily a change in what versions of gnucash are available but a change in what version actually installed. Some people like to always install the newest and "best" while others tend to wait till all warts removed from a new version (aka "avoid the bleeding edge).

c) Theoretically when playing against a computer program you should win by a large amount or lose by a small amount. If a program predicts that playing normally it will lose by a small amount its only rational choices are "resign" or "take any chance to pull the game out". A human player might take solace from a loss "at least it was close" but we humans aren't entirely rational (computers have no emotions at all). So going by margin of victory can be very misleading. Don't assume "if I am losing by only ten points I will win if I take one more handicap stone". Not unlike when a very strong player is teaching us.

Option A for 200, Alex!

People here are comparing Gnugo's play today (i.e. the current version) to its play from x years ago (5 or so in my case since last time I played it.) At that time I could give it 9 stones easily and win by killing most of the board, but now 5 or 6 looks more reasonable.

[jts]

Does gnugo get stronger when it runs on faster equipment? Gnugo didn't seem to be very strong on facebook, for example, and I assumed that was because whoever provides that service couldn't devote too much processing power to it.

If that's the case, that would also explain gnugo getting stronger over the past 5 years; presumably the computers it runs on have been getting better.

[karaklis]

I thought this "win by large margin, lose by small margin" is only a matter of MC based engines. But GnuGo is not an MC engine as far as I know.

Your strength strongly depends on the time settings. Engines like GnuGo are stronger at blitz, maybe 3-4 stones. I use GnuGo 3.8 (the latest version) and play it at three stones with about 15 minutes per player.

[Mike Novack]

a) Five years! That is a very long time in the state of the art of go playing programs. Yes, gnugo has gotten susbtantially better over the last five years (and so have all the others that are still "in the game")

b) Not exactly. I was describing "logical play strategy" (it applies to us humans too). It's automatic that programs using an MCTS evaluator will behave this way. But even an AI evaluator program should behave this way; some logic based on "am I ahead or behind?". Many of the AI go playing programs have parameters controlling their style of play and this affects which move their AI "evaluator" might choose (from all possible moves with some "go purpose). Again, I wasn't saying that gnugo had code in place to do this, just that you shouldn't be surprised if it did behave this way because it should.

[liquido]

GnuGo 3.9.1 was released a short time ago. I haven't measured so see if there is a noticeable difference, but it's quite likely. Also version 3.8, which has been out for some time, is quite a bit stronger than 3.7 from what I've heard.

Also, MC programs would be "win by a small margin or resign", unless the author has specifically tried to address this "issue."

[topazg]

That would make sense. The improvement was within the last few months, and I'd say it's more like KGS 4 or possibly 5 kyu now