Simple Sure Win Strategy for White Human Player vs GnuGo 3.8

[GloFish]

I found an amazingly simple sure win strategy for a white human player versus GnuGo 3.8, without handicap. So far I always won when applying an ultimate cosmic style, playing a shoulder hit to every black move in the opening. If black answers with keima, play keima too. If black attaches, just extend - or play hane, if black continues to push from behind. GnuGo is virtually always happy taking all 4 corners starting on 3-3, 3-4, 4-4 or 3-5 points. If Gnugo goes for a more fancy move like 5-5, still apply the shoulder-hit on 6-6 - and invade the corner later. There is no real middle-game, and the end-game is quite easy. Just play all the sente moves once you get sente, and calmly answer all peeps. I'm dan level player, but I guess single digit kyu players should be able to handle it as well. At least as long as GnuGo doesn't decide to go for an invasion - which is totally hopeless though, if you know some basics of attack and defense, and GnuGo gives up on the invasion after about 5-10 turns. I get final scores of about 10-60 points for white, komi not yet included.

[tj86430]

Can you post some example games?

[gaius]

Lo and behold the weakness of non-MCTS go software!

Still, nice find

[topazg]

To be honest, anyone who is looking for a strategy to win encapsulates why most people suffer from playing bots.

The goal isn't to win, it's to play good go. If you aren't playing each move thinking "is this the best move I can play", don't play it. Just because you know the bot will respond wrongly doesn't make a move right.

[RobertJasiek]

It depends on what one wants:
- improve oneself
- win a lot
- research program weakness

[Toge]

To be honest, anyone who is looking for a strategy to win encapsulates why most people suffer from playing bots.

- Learning Go teaches moves and sequences that often lead to victory against other human opponents. Playing shoulder hit on certain conditions is better than never playing shoulder hit. Now we know that GnuGo has mistaken belief that the way it is playing is good. If Go on 9x9 board is ever solved, we'll know if our gut instinct to start at the center was good or not.

[Mike Novack]

I found an amazingly simple sure win strategy for a white human player versus GnuGo 3.8, without handicap. So far I always won ......... I get final scores of about 10-60 points for white, komi not yet included.

Silly question? You say that you are 1 dan? Do you have some reason to suppose that you shouldn't be able to beat GnuGo 3.8 every time at no handicap? Whether you play the way you have described here or with some different strategy?

In other words, how strong do you believe GnuGo 3.8 is supposed to be? If GnuGo is several stones weaker than you are you need not look further for why you are winning all the time. What you describe here is of interest; but mainly to the developers of GnuGo's AI algorithm (as a specific case of where they may need to do more work).

What you should be doing if you want to research if what you think you have found holds true is to look up how strong GnuGo 3.8 is supposed to be and then allow it as many handicap stones as you would allow a human opponent of that strength to make a fair game of it. Then see if your strategy works better than other strategies. Remember to consider only the precentage of games won or lost because that is how go is scored, not by margin of the victories.

[karaklis]

GnuGo is not stronger than 10k. It would be more interesting (and more appropriate for a dan player) to find a way to win against GnuGo with nine stones.

[amnal]

GnuGo is not stronger than 10k. It would be more interesting (and more appropriate for a dan player) to find a way to win against GnuGo with nine stones.

I always thought it was a bit higher than that. I can give it 7 stones fairly easily, but 9 is hard. Fun, though...although its rank is consistent, gnugo does not play like a human at all, which is interesting.

[RobertJasiek]

and more appropriate for a dan player

I appreciate everybody's research results regardless of whether he is kyu or dan. If kyus want their share in pride, then they must make discoveries before the dans do. The kyus have not found it, so all the pride goes to GloFish.

[Toge]

I appreciate everybody's research results regardless of whether he is kyu or dan. If kyus want their share in pride, then they must make discoveries before the dans do. The kyus have not found it, so all the pride goes to GloFish.

- Pride is not a virtue.

[amnal]

Why not?

[Toge]

It's two different things to be good and to think that you're good. Pride has many negative implications. Idea that other people are less than me. They have nothing to teach me. Belief that what is good has a source in individual itself. That is directly discrediting people around you who no doubt contribute to your success. Many people also use pride to give impression that they are of higher social status.

I much prefer humility.

[RobertJasiek]

Pride (and being admired) motivates research discoveries (or winning tournaments) much more than humility, and pride together with joy (about doing research) more than joy (or a mere interest in understanding) alone. With humility instead of pride, science and go theory would be at much lower levels so far because speed of discoveries would be much slower. Regardless of whether pride is not considered a religious virtue, it is an important motivator. Demanding humility instead of pride and showing ignorance instead of admiration can be demotivating instead of motivating. We do not live in an "I have found the wheel!" - "Who cares?" - world but in an "I have found usage for fire!" - "Great!" - world!

[gaius]

^ Robert, I don't know about Germany, but here in the Netherlands, both fire and the wheel are not considered very cutting-edge technologies any more...