Simple Sure Win Strategy for White Human Player vs GnuGo 3.8

lorill · Post by **lorill** » Sat Jan 29, 2011 6:39 am

Sorry, I don't feel the need to continue the discussion. Try to find some real games with meat&bones players. There are 8 stones difference games played everyday (check out the ASR League for instance), go look for a few of them.

daal · Post by **daal** » Sat Jan 29, 2011 10:58 am

Mike Novack wrote:We need to look at examples of human 8k's trying this "easy win" strategy against gnugo 3.8 . We may find that enough errors will be made in the josekis and sealing off and use of aji that the outcome is different. Conversely we may find that gnugo has a lower effective rating. But looking at games where the human player has about an 8 stone edge tells us little.

I'm 6k and I just beat gnugo following the above strategy and otherwise playing mindlessly. I am sure that it would take less than 1 hour to teach a 16k to beat it.

GloFish · Post by **GloFish** » Sun Jan 30, 2011 12:09 am

daal wrote:I'm 6k and I just beat gnugo following the above strategy and otherwise playing mindlessly. I am sure that it would take less than 1 hour to teach a 16k to beat it.

Cool, thank you for your report. I'm looking forward to reports of even weaker players, how they could deal with that strategy. My statement that 1 digit kyu's can apply the strategy was only a guess after all, I wonder if it holds.

Mike Novack wrote:We learn from what succeeds, not from what fails. We don't/won't learn to play properly if our only game experience is playing even games against players that much stronger than ourselves doing their utmost against us

Sorry Mike, but I think you're totally wrong here. From even games against much stronger players you learn most, cause you can see what moves your opponent applies to crush you, and then try them in the next games. And you also learn from your failures. When a move doesn't work, you learn that it doesn't work and try another one next time, until you found a move that works. And once again, you learn most effectively when playing against a much stronger player, cause you won't be left with any illusions. If your move doesn't work, the stronger player will punish it and thus support your learning. I'm well aware that this kind of learning might be demotivating for some people - nevertheless it is the most effective way to become stronger. I've always challenged stronger players at Go congresses as often as I could - not to win, but to see Go games at high level and learn from them - and I was well rewarded!

hyperpape · Post by **hyperpape** » Sun Jan 30, 2011 4:48 am

Yeah, I'll second those who say an 8k human would not play like this. They might make some of the same initial mistakes, but they would invade.

I'm curious to see how weak a player would need to be before they couldn't apply this strategy. I'm not so sure a 16k wouldn't flub it--just push a little too much, end up with too few liberties, and you're in trouble.

Mike Novack · Post by **Mike Novack** » Mon Jan 31, 2011 5:09 am

GloFish wrote:
daal wrote:
Mike Novack wrote:We learn from what succeeds, not from what fails. We don't/won't learn to play properly if our only game experience is playing even games against players that much stronger than ourselves doing their utmost against us

Sorry Mike, but I think you're totally wrong here. From even games against much stronger players you learn most, cause you can see what moves your opponent applies to crush you, and then try them in the next games. And you also learn from your failures. When a move doesn't work, you learn that it doesn't work and try another one next time, until you found a move that works. And once again, you learn most effectively when playing against a much stronger player, cause you won't be left with any illusions. If your move doesn't work, the stronger player will punish it and thus support your learning.

Not reading what I wrote?

Unless the much stronger player is playing a "teaching game" with you, won't learn that stuff. I was not suggesting that one didn't learn best by playing against a much stronger player. I agree entirely that this disparity in strength will quickly let you see what you are doing wrong, immeidately punish mistakes. It's the "even game" part that I am disagreeing with. To learn fastest what you want is for that much weaker player to have a slightly inadequate handicap.

In other words, I am saying that an 8k won't learn that that much playing a 1d even but would if playing with a 5-6 stone handicap.

jts · Post by **jts** » Mon Jan 31, 2011 6:55 am

Mike Novack wrote:
In other words, I am saying that an 8k won't learn that that much playing a 1d even but would if playing with a 5-6 stone handicap.

An "even" game there would technically be a 8 stone handicap, no? But regardless. I've improved a lot while playing even games with players 5-9 stones stronger than I was. Reviewing even games against stronger players makes your mistakes much more stark (because up until a certain point, you were doing better).

Mike Novack · Post by **Mike Novack** » Mon Jan 31, 2011 10:24 am

But that is not how the term "even game" has been used in this discussion (remember the examples?)

Jts, you meant "even" in the different sense (with handicap) and I couldn't agree with you more. Playing even games (no handicap) against players as weak as you are will let you get away with too many mistakes. Playing with the game slightly uneven (slightly inadequate handicap) will highlight what you are doing wrong. Frankly, it's no always necessary to go over the game to see. At least in my case usually immediately obvious "darn; I needed to make a defensive move and that was a defensive move that retained sente because with that attacking group safe the attack (that otherwise falls to a counter attack) would win through" --- I give that example because missing these is a current problem of mine.

But I don't think 8-9 stone games best. An opponent several stones stonger will be enough stronger and with the initial board more open (fewer handicap stones) the game closer to an "even game" (no handicap). Of course you can learn from mistakes you are making that an opponent 8-9 ranks higher can spot and punish but that might succeed against an opponent just 4-5 ranks higher. But I say "walk before you run". It's the mistakes typical of your current level that a player 4-5 ranks higher doesn't make you should learn first. Wait till later to learn the things you are doing wrong that a player 4-5 ranks stronger also does wrong.

This being the "computer go" section I think this is very important for those of us who have* to use the computer as an opponent. Just my opinion, but the "bad habits" problem should be far less of a problem if the computer is "peculiar" at its own level of play. Those far less likely to manifest if the playing strength of the program is adjusted to 3-6 stones stronger than you are.

* I get to go to the go club once a week and play against humans there and this time of year that's "weather permitting" (snow either the day the club meets or the day before and too sore/tired after digging out to drive any distance and then stay out late) I can play a game or two a day against the machine with a program enough stronger than I am so I can learn from it.

jts · Post by **jts** » Mon Jan 31, 2011 11:34 am

Sorry, I (stupidly) switched from one sense of even to another in two consecutive sentences. I play no-handicap, 6.5 komi games with stronger players as part of ASR, and I've been learning a lot from it.

Mike Novack · Post by **Mike Novack** » Tue Feb 01, 2011 7:06 am

A bit of research (gnugo developers correct me if I am wrong)

Since no users of gnogo responded to my semi questions I went to their site to see what I could determine.

a) Deterministic only --- I could find nothing referring to a mode where probablity is involved with move selection. That means gnugo will always respond exactly the same way in the same situation and a weaker player could "learn a game" that defeats it. In other words, it represents a more general weakness than the specific "bad game" example that started this thread.

b) Center influence ----- This is adjustable, at least selectable even if off/on vs a continuously adjustable parameter. But hasn't been made available as a run time parameter so you need to be able to recompile the program ot turn on "cosmic".
The site says that with this on beats gnugo with this off but weaker against human players. That would contradict what has been suggested here that a weaker player could easily "learn" to use gnugo's passivity. We really need to here from players weaker than gnugo is described to be that they are able to do so. Sorry, but that stronger players can do it means nothing as they should defeat gnugo in any case.

Were I on the development team I would ..........
1) Make the "cosmic" option available as a runtime option. Even if off/on. Could be even more interesting to make the "off" vs "on" a matter of probablility (parameter changes the odds whether gnugo will play this game in an "territorial" or "influence" style.

2) Introduce probablity in move selection. Bad to weaken the evaluator too much so I would leave it deterministic if the top move evaluates more than the second best by a certain amount, but this can be almost automatic if the choice between the two top moves is proportional to their metrics. This would make it much harder for a weaker human opponent to "learn a game that defeats gnugo all the time".

However those two suggestion would be only if continuing not to use MCTS as the "evaluator". As things stand now there are two parts to a program like gnugo. One is the AI that has processes that find plausible moves (moves that might be good for some go reason) and the other is an AI that is the "evaluator", tries to determine which of these is the best/most important move at the moment given the current state of the game. But MCTS can be used for the "evaluator" instead.

daal · Post by **daal** » Tue Feb 01, 2011 8:28 am

Mike Novack wrote:In other words, it represents a more general weakness than the specific "bad game" example that started this thread.

Huh? The example that started this thread was not a specific "bad game," but rather pointed exactly to a "general weakness."

Sorry, but that stronger players can do it means nothing as they should defeat gnugo in any case.

While you go on to make some of your own valid points, particularly about how GnuGo works and how to improve it, you seem to be missing the point of this thread, which is that a simple mechanical strategy tricks GnuGo into playing a losing strategy. It is irrelevant that a stronger player would win anyway. When I tested the OP's strategy, I did not attempt to win, I simply attempted to employ the theory, and GnuGo played exactly as the OP predicted. That's the point.

daniel_the_smith · Post by **daniel_the_smith** » Tue Feb 01, 2011 8:41 am

Gnu Go uses a random seed, it is deterministic given that seed but it changes the seed every time it is launched (unless you request a specific one be used). I believe it uses that to choose between opening moves, I'm not sure what else. I think there's sufficient variation to make it impractical to memorize a game tree that will always beat it.

Mike Novack · Post by **Mike Novack** » Tue Feb 01, 2011 1:52 pm

Is the interest in how a computer program loses when badly overmatched of interest? (that it loses in one way rather than loses in another).

This isn't specific to gnucash and the other pure "go knowledge" engines. Try playing one of the MCTS engines under game conditions where the program is badly overmatched (and especially if you refuse it permission to resign). You will conclude "this program is very weak; makes terrible moves I wouldn't expect from a human player many ranks weaker than the program supposedly is". Will conclude that it loses differently than a human player does.

Under those conditions.

Asking too much? At the current time all efforts with these programs is to get them to play as strongly as they can under game conditions of comparative equality. I don't think any development effort has gone into their behavior when badly overmatched.

It is of interest (and was suggested) that this "simple strategy" could be successfully applied by a player much weaker than the program (its supposed strength). That remains to be demonstrated. Sorry, but stronger players saying "I wasn't really trying" isn't a demonstration of that since being able to see "coordination of influence" is begins to be gained at about this level of play. Look at commented games between say 8ks and there will be plenty discussing "wrong direction" and "not seeing coordination". This "simple strategy" isn't so simple for a 10k to properly apply.

snorri · Post by **snorri** » Tue Feb 01, 2011 3:29 pm

Mike Novack wrote:A bit of research (gnugo developers correct me if I am wrong)
Were I on the development team I would ..........
1) Make the "cosmic" option available as a runtime option. Even if off/on. Could be even more interesting to make the "off" vs "on" a matter of probablility (parameter changes the odds whether gnugo will play this game in an "territorial" or "influence" style.

Um, in GNUGO 3.8 there is a --cosmic-gnugo option you can pass on the command line at runtime. Have you tried that?

Edit: I just tried it and it plays a very different game. However, the strategy still works, maybe just less dramatically. The monkey jumps at moves 21 and 23 were particularly surprising to me, but to play white's trick strategy, one has to find some local response because you want to wait for black to take another corner.

jts · Post by **jts** » Tue Feb 01, 2011 3:54 pm

Mike Novack wrote:Is the interest in how a computer program loses when badly overmatched of interest? (that it loses in one way rather than loses in another).

I really think you're missing the point. Why don't you play a game against gnugo using this simple set of instructions, and see what happens? I think that will give you the flavor for what we're talking about. I tried a day or two ago and beat gnugo by 37, with absolutely no reading, just mechanical application of the strategy. Glofish may be 1d, but the instructions he came up with are not 1d instructions.

Mike Novack · Post by **Mike Novack** » Tue Feb 01, 2011 7:20 pm

jts wrote: Glofish may be 1d, but the instructions he came up with are not 1d instructions.

Who suggested that they were? (1 dan instructions)

But are you willing to agree that they are perhaps 6-8 kyu instructions? Gnugo isn't supposed to be anything like as strong as 6 kyu.

The instructions are (in effect) to coordinate one's separate positions so that they have a combined effect, that gnugo has a weakness in that it will not attempt to contest this. But isn't that sort of coordination precisely the sort of thing human players as weak as gnugo have trouble with?

To demonstrate that this is a method that human players significantly weaker than gnugo could use to easily defeat gnugo need to have games between players of those strengths an gnugo. Since there are a number of bots using gnugo, all we should have to do is wait and see if there is a noticable decline in the rankings of these bots.

Life In 19x19

Simple Sure Win Strategy for White Human Player vs GnuGo 3.8

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo

Re: Simple Sure Win Strategy for White Human Player vs GnuGo