Playing computers finally good for your game?

[Oroth]

With recent advances in Go playing programs, is playing the computer finally becoming a viable means of improvement - for kyu players at least? People often say that the best way to improve is to play a better player and having Many Faces or zen19 would be like having a dan player continually on hand. Conventional wisdom has always been that playing computer programs is bad for progress, partly because computers have always been weak, but also because they play a lot of strange moves. Computers are no longer so weak, and the stronger ones seem to make less strange moves. So maybe playing lots of games against the computer is a good way for kyu players to improve, or at least ddks. I can't help but think a player would gain more from playing a program someone ten stones stronger than playing someone their own rank, even if the program does play some funny moves occasionally.

[freegame]

Of course you can play against the computer. Of course you will get better from it if you are weaker than the bot. This has not changed recently, just the nr of people who can profit from it has increased because the bots got stronger. The quality of the games bots play has not improved I think. Bots still play like bots and not like humans (there really is a "visible" difference)

If there is no human player to play against playing against a bot is a good alternative. Just don't play only against bots and expect your results to be similar against other humans. If you play some games against bots and some against humans that is not "bad" in my opinion

I recently tried out Many Faces of Go v12.
the bot plays indeed stronger that the one in the previous version but (on 19*19) it's no EGF dan bot. maybe with luck it gets a weak 1d on KGS but i think it's real rank on KGS is somewhere around 1-2 kyu (Just my oppinion based on playing a dozen or so games against it with default time settings and a reasonably high performance home desktop computer.

on a 9*9 board it's quite a challenge for me to beat it but I'm in general not very strong on a 9*9 board.

[topazg]

Computers being strong doesn't mean computers being good to learn from. My experience with Monte Carlo based bots (most of the strong ones) is they play bizarre moves. They tenuki when they shouldn't, they completely implode when they're behind, the play crazy fusekis and concentrate on reading power overpowering a human in the midgame.

I think bots have got stronger, but less appropriate to learn from in my experiences against MFoG and Fuego.

[CarlJung]

They tenuki when they shouldn't

Yes, they leave simple corner shapes to die in ways that no human would dream of. This makes it very difficult to learn from them since they don't show the proper way to play. At least not on 19x19. on smaller boards it's another matter.

[Cassandra]

In Chess, programms were first used as analysing tool, even by Grand Masters, long before the programms reached Grand Master's playing strength (and turned into a challenging opponent).

I think, this will prove true in Go.

[kirkmc]

In Chess, programms were first used as analysing tool, even by Grand Masters, long before the programms reached Grand Master's playing strength (and turned into a challenging opponent).

I think, this will prove true in Go.

I think you're talking about using them as databases, not as playing programs. It's very possible that pro go players already do this.

[Cassandra]

I think you're talking about using them as databases, not as playing programs. It's very possible that pro go players already do this.

I'm not talking about databases.

For example:

If the Go playing program has an output function for what it thinks may be the winning propability for every (human played) move, it may be interesting to have a look at the discontinuities in the graph.

If the Go playing program has an output function for what it thinks may be the winning propability for every move the program has been "thinking" about (= "find the next move" in the game), it may be interesting to compare moves with an approximately equally value.

[xed_over]

If the Go playing program has an output function for what it thinks may be the winning propability for every move the program has been "thinking" about (= "find the next move" in the game), it may be interesting to compare moves with an approximately equally value.

yeah, I used gnugo for that to review my own games, at least until I got to about the same level as gnugo. It would give weighted multiple possibilities for moves I had not even considered.

[karaklis]

I agree with topazg that MC engines behave very strangely sometimes. Once I checked out Fuego, and even though I am clearly weaker (around 6 stones), I was able to crush it. I had a similar experience with Leela on a 13x13 board. Both programs made very unreasonable and strange moves.

As for the traditional engines like GnuGo and Aya, I was told not to play them any more once you can beat them. But what about handicap matches? At the moment I am trying out to play Aya in fast games, giving the program some handicap stones. That is because I have two special weaknesses: When playing fast games, I am around 3-4 stones weaker. And I am weak at giving (and receiving) handi stones. Is there something to object against using those programs for this purpose?

[dfan]

I'm kind of unfortunately addicted to playing the Aya bots on KGS. They're supposedly around my level but I seem to be able to beat them around 80% of the time; I may just be good at adjusting to them. It is probably not so great for my go overall, since it doesn't play much like any human does, but there are a couple of skills Aya tends to test well (I don't know so much about other bots):

- It tenukis a lot, so it forces to really think about whether your previous move was sente, and how to prove that that move was more valuable that its tenuki.

- It likes to make big moyos, so you have to be on guard to not let it complete them. On the other hand, it is pretty bad about letting you worm through them once you poke your way in, in a way that reminds me more of 10k games than 4k ones.

[flOvermind]

I think the problem with the bots is not their strength, but the consistency in their mistakes. The current bots may play around 1d level, but they will always make the same type of mistakes.

Let's say you're 5k, and the bot is 1d. You play the bot, and soon you are able to beat it. Does that mean you're 1d now? No, you just learned how to play against that particular bot, punishing its mistakes.

It doesn't really matter how strong the bot gets, this problem will always apply as long as the bots make significant mistakes. It would only work if you know that the bot is stronger than you in (nearly) every aspect of the game. As it is now, the bots are reasonably strong on average, but ridiculously weak in some aspects of the game. Playing too much against bots (or should I say not enough against humans?) will just teach you to exploit that, or even expect it, leading to bad habits.

[zinger]

I think that bots will work their way into top level go the same way they did in chess: as endgame analyzers. The endgame is much more "computable" than the opening or middlegame.

Which leads to a really interesting weakness in the bots that topazg mentioned - "they completely implode when they're behind". Already the endgame is their strength; therefore they should play out endings where they are behind, and will often win. But instead they just assume their opponent is infallible and effectively give up.

An interesting comment on chess, from a friend paraphrasing Dvoretsky:

Basically he says that—while people like Kramnik and Anand, Ivanchuk and Leko, etc. understand chess at the highest level, like their predecessors—a lot of today's leading players actually aren't all that good. (Yes, better than we are, but that's not the relevant standard.) Outside of tactics and computer-friendly opening variations of course. They outsource strategic judgment and creativity to their engines; and they don't really know what deep analysis of an endgame even means. Take away their silicon toys, and they'd be eaten for breakfast in the classic Soviet championships.

I truly hope, that computer endgame crunching never does this to top level go.

[CarlJung]

Which leads to a really interesting weakness in the bots that topazg mentioned - "they completely implode when they're behind". Already the endgame is their strength; therefore they should play out endings where they are behind, and will often win. But instead they just assume their opponent is infallible and effectively give up.

They don't give up. They play unreasonable moves. Unreasonable to the point that you have to be 30k in order to miss the obvious answer to their plays.

[quantumf]

Which leads to a really interesting weakness in the bots that topazg mentioned - "they completely implode when they're behind". Already the endgame is their strength; therefore they should play out endings where they are behind, and will often win. But instead they just assume their opponent is infallible and effectively give up.

They don't give up. They play unreasonable moves. Unreasonable to the point that you have to be 30k in order to miss the obvious answer to their plays.

I think the point was that by playing silly moves they've effectively given up. Instead, zinger argues, they should regard their perfect endgame play as being their weapon to recover from situations where they are behind.

[dfan]

An interesting comment on chess, from a friend paraphrasing Dvoretsky:

Basically he says that—while people like Kramnik and Anand, Ivanchuk and Leko, etc. understand chess at the highest level, like their predecessors—a lot of today's leading players actually aren't all that good. (Yes, better than we are, but that's not the relevant standard.) Outside of tactics and computer-friendly opening variations of course. They outsource strategic judgment and creativity to their engines; and they don't really know what deep analysis of an endgame even means. Take away their silicon toys, and they'd be eaten for breakfast in the classic Soviet championships.

This is an interesting take, but you would find many, many people who would disagree with it.

(The one part I do agree with is that the ability to do deep analysis of endgames has decreased. This is due to shorter time controls and the death of overnight adjournments, though, not to computers.)