Life In 19x19

I've recently got a bot running on KGS to see how it did, running GnuGo 3.8 - its playing strength from the games it has played so far puts it in the range of 5-10k at the moment.

karaklis wrote: When I tried that, I got outplayed all the time, so that's why I discovered that it is weak at moyos. I'd like to experiment with new tactics, so maybe you'd like to share your ideas? This is the way I usually beat GnuGo.

This is an interesting point, apparently there is an experimental setting for GnuGo to try and play with a focus on moyo but since it would make the engine weaker it is not enabled by default. I may recompile it and see what effect it has at some point.

karaklis wrote:

Wildclaw wrote:

- Stay away from local fights
- Try to establish big moyos

Disagree on this. In fact, starting fights is one way to get ahead vs GnuGo, because it tends to back off, tenuki or fail to consider the strategic consequences in fights.

As for establishing big moyos. While there is some truth in that, it is just one way of exploiting the passiveness of GnuGo, and generally not the best way in my opinion.

Interesting. You are some 6 stones stronger than me, maybe this is the reason you are stronger at local fights than the program. When I tried that, I got outplayed all the time, so that's why I discovered that it is weak at moyos. I'd like to experiment with new tactics, so maybe you'd like to share your ideas?
This is the way I usually beat GnuGo.

Here's a fun game I played the other day against GnuGo with 6 stones - you can decide for yourself, but I agree that it feels like it can't handle complicated fights - particularly kos, which it seems to make a repeatedly bad job of.

Interesting match. I discovered that you often tried to wedge through an ikken tobi. I thought that this often causes bad shape...

karaklis wrote:Interesting match. I discovered that you often tried to wedge through an ikken tobi. I thought that this often causes bad shape...

It does, but it's also in quite a few josekis. It generally leads to lots of split groups - I find GnuGo plays passively slow, and then when you create a cut fight like wedging through an ikken tobi, it goes a bit crazy instead of solidly connecting the important bits, and then picks odd times to tenuki. Time after time it can be ripped off in this way.

Normally the proper shape invasion of the 4th line 3 point jump is in the middle, also on the 4th line, but I feel this illustrates more GnuGo's failure to handle complicated fights well.

I see this thread is a few years old, but I'm sure that GNU-Go has only gotten better since then. I would like to offer my thoughts on GNU-Go, from a novice/beginner's standpoint. Perhaps most of you are very good at this game, but I would hope you would be able to see my thoughts from back when you all were not so good.

GNU-Go is far too difficult. This is absolutely a turn off for new players. There are two main issues here. first, GNU-Go almost always plays a move that will gain the most territory possible. A novice is not capable of moves like these. Second, GNU-Go will leave you one move away from capture in almost all situations. For a novice, it's hard to discern. While GNU-Go begins to play elsewhere, gaining strength and territory, the novice continues to play for a capture only to have GNU-Go drive the dagger in deep at the last moment, breaking the line.

It's saddening, as a novice/beginner, to even use GNU-Go and instead of making me a better player it only makes me never want to use it again, and further, give up attempting to learn any further.

t1m1976 wrote:GNU-Go almost always plays a move that will gain the most territory possible.

IMHO that's not possible, otherwise GnuGo would not stick in the 6-8 k range

t1m1976 wrote:Second, GNU-Go will leave you one move away from capture in almost all situations. For a novice, it's hard to discern.

However that is one important principle in Go: If you have safely captured a group, you should not continue killing it immediately. Instead, it is much better to play in a region, which is not yet decided. It is a common mistake by beginners to kill captured groups completely (maybe in order to make things on the board more clearly), however this wastes much time, while the opponent can play in unsettled areas and get much profit there.

t1m1976 wrote:While GNU-Go begins to play elsewhere, gaining strength and territory, the novice continues to play for a capture only to have GNU-Go drive the dagger in deep at the last moment, breaking the line.

Yes, for the above reasons, GnuGo only completely kills, if it is absolutely required (e. g. to get rid of an atari).

If you post one of your games against GnuGO (or alternatively, have a look at http://gtl.xmp.net) you could get some good first hints by stronger players. Also I recommend to play rather human players (e. g. in the double-digit-kyu range on KGS) than bots.

If you get stronger after some time, IMHO playing GnuGo is a bit "dangerous": It tends to tenuki too early in local fights, which it can-not calculate completely, so it is comparable easy to establish an invasion in an area, where it normally shouldn't work. If you switch back to human players, your play might have become too careless, and you will lose too many groups.

t1m1976 wrote:...... While GNU-Go begins to play elsewhere, gaining strength and territory, the novice continues to play for a capture only to have GNU-Go drive the dagger in deep at the last moment, breaking the line.

It's saddening, as a novice/beginner, to even use GNU-Go and instead of making me a better player it only makes me never want to use it again, and further, give up attempting to learn any further........

a) You presumably have the bot set to be playing too much stronger than yourself if you are never winning a game.

b) You have the bot set not enough stronger than yourself to learn from if you are winning more often than every once in a while.

c) If YOU are losing by making the same mistake over an over you need to ask yourself "why am I not learning how to deal with this situation?". For the particular situation you describe (misreading life and death) you need to try doing some life and death problems.

d) Take another look at "a" and "b" and realize that this is not exactly a machine vs human opponent issue. If you are unsatisfied playing except against opponents you can beat 50% of the time your rate of learning will be very slow. You want to learn from opponents who make fewer mistakes than you do. At this low level, the non-MCTS programs probably better as teachers as they are more consistent (but if provided by an AI, select the "randomize" option -- slightly weaker but will not always make the same move).

3) Depending on you level (and you politics) you might have more options for free as in free beer software. If you are in the 20's or worse, the free demo version of MFOG might be useful. This is not "free software" but the 18kyu only demo version is free for download and includes some of the other features (the beginner level problems, etc.). I haven't tried this limited version (I have a MFOG license so for me, not limited) so I don't know whether it includes all the other features but you should at least check it out for that -- problems, fuseki, options like "tell me why" (why did it make the move that it did). However once you are ~20kyu or better the free version of MFOG probably no longer useful.

t1m1976 wrote:I see this thread is a few years old, but I'm sure that GNU-Go has only gotten better since then. I would like to offer my thoughts on GNU-Go, from a novice/beginner's standpoint. Perhaps most of you are very good at this game, but I would hope you would be able to see my thoughts from back when you all were not so good.

GNU-Go is far too difficult. This is absolutely a turn off for new players. There are two main issues here. first, GNU-Go almost always plays a move that will gain the most territory possible. A novice is not capable of moves like these. Second, GNU-Go will leave you one move away from capture in almost all situations. For a novice, it's hard to discern. While GNU-Go begins to play elsewhere, gaining strength and territory, the novice continues to play for a capture only to have GNU-Go drive the dagger in deep at the last moment, breaking the line.

It's saddening, as a novice/beginner, to even use GNU-Go and instead of making me a better player it only makes me never want to use it again, and further, give up attempting to learn any further.

Hi T1m. I take it you're a new player? Welcome to the game!

In general in Go, the correct way for beginners to get games that they have a good chance of winning is not to ask the other player to make poor moves, but instead to start with stones already on the board. This handicap system isn't just for beginners, either... in three hours I'll show up to my go club and I might be taking two stones from our strongest player or giving six stones to one of our newer players.

I would suggest that you start by playing, not just GnuGo, but a much stronger AI called Fuego, on a 9x9 board with 9 handicap stones. Should you win this easily? Oh yes you should. Then every time you win you can take away a stone. Where will the game start to get really challenging? 4H, 3H? It's hard to say. But the point is, you should be able to set up a game where you are seeing very strong moves that surprise and delight you, while also getting a game where if you play to the best of your ability you win and if you slip up you lose.

The 9x9 board is to make it easier for you to keep track of everything that's going on, to see the connection between opening, middle game, and endgame more quickly, and so that each handicap stone has a bigger effect.

By the way, it's rather impressive that you've noticed the biggest differences between the way GnuGo plays and the way beginners play all on your own without any coaching! Not making unnecessary moves, and ignoring your opponent's moves to make a lot of territory elsewhere - both huge steps.

I feel like I should respond because I started this thread way back in 2010. I can't remember this post, but it is amusing to read it again .

Having bots to play is wonderfully convenient, but I can't really attribute any significant part of my improvement to bots. Rather, playing people and reviewing the games (especially with a stronger player, but even alone) has by far been the most beneficial method of study for me, along with doing go problems.

As for playing bots, I echo the advice given by others in this thread. And don't be reluctant to take a handicap.

EDIT: And it looks like I was about 14k at the time. I've gone back and looked at a few game records from around that time...pretty amusing to me now.

Since I have struggled with GnuGo and other bots long time ago, here are my suggestions:

When I was a beginner I hated to play with handicap stones, but in retrospect I see that this was wrong. So, start over with 9 handi stones. If you cannot beat the program, start with an easier program. For a long time I liked to play with Aya 6.34 (as far as I know this is still available for download). Its strength should be somewhere between 12 and 15k, so as a beginner you can start with taking 9 stones. Once you have beaten the engine, reduce the handicap. When you are able to beat Aya without handi stones, you should be able to play GnuGo with five stones. Then continue with the described procedure until you can beat GnuGo, or you can give it stones if you like.

It should not be too much of a surprise. GNUGo and other have been improved greatly. There is a juicy reward for those AI engineers that can produce a Go-playing software capable of beating a 9p in an even game. A tall order, of course, but it is a goal many AI engineers strive for. At the very least it would be a new pinnacle in their art.

I'm pretty sure those geeks get really pissed off when some professional Go player tells them their latest software is garbage. It only spurs them on to revenge. Hence the increasing strength of Go-playing software.

I usually never win against GNUGo. It plays at around 6k to 3k in strength, so unless I have a handicap 9 out of 10 times I lose. If I do win against it, it's because I play aggressively by attempting to develop quickly all over the board in the opening. Of course, as the number of possible moves is reduced, the software's powers of calculation become more apparent.

Someday I'll be able to beat GNUGo in an even game by 100 pts. Until then, I've no choice but to replay game records, do Go problems, and play actual games - of which I have not done much lately except on turn-based servers For any novice learning Go for the first time, I recommend the above regimen. A couple of professional game records daily - to get a feel for what a game of Go is like, minimum of 15 minutes of problem solving daily - to train your brain to analyze possible moves in any board position, and playing at least 2 games a week, preferably against a human - to apply what one has learned. (I would play Gobots for fun.) Depending on one's mood the time spent doing each item can be increased. All the better to help one improve; getting good at Go is hard work. No need to stress it out, though. Even pros had to study long hours and have lots of patience to get that pro certificate. Just take it easy and put in a little bit daily. Slow and steady wins the race.

tekesta wrote:GnuGoBot38 and Bayi, another SDK-level Gobot, play at middle to high SDK level on KGS, so unless I take handi I rarely win against either and the losses are uglier when I do not put in enough analysis time.

As for Bayi it doesn't accept handi stones and only plays free games. It sometimes plays strange games. I have seen it beat 1k players, and in the very next game lose to a 9k player. For those who are interested, Bayi cannot be found in the computer room, but plays in the Indonesian room ('bayi' is Indonesian and means "baby').

Playing a stronger bot is frustrating for the same reason playing a stronger human is frustrating: Go is a frustrating game. As a beginner I probably should have stayed on 13x13 for longer, reduce the complexity and get better at some aspects of the game. There really is a lot to consider when you start and you get lost.

One thing I liked about gnugo however, is that it played the same tesuji, cutting, killing and playing some strong yose moves I did not know, over and over. And of course, being a bot, it didn't mind my undoing a few moves to try and correct my mistakes and check the score estimate difference. I found that helpful. But I didn't learn about playing an aggressive opponent, so it was not balanced.

A stronger MTC program with classic fuseki play, like MFoG12 would probably be best these days. Having a 3/4 dan opponent at your beck and call, even if it has flaws that can be exploited (like any 3d, otherwise it'd be a pro) and play it on 9 stones handi or whatever, is bound to be helpful.

Dudoso

Dudoso wrote:
A stronger MTC program with classic fuseki play, like MFoG12 would probably be best these days. Having a 3/4 dan opponent at your beck and call, even if it has flaws that can be exploited (like any 3d, otherwise it'd be a pro) and play it on 9 stones handi or whatever, is bound to be helpful.

Dudoso

1st of all, to have a 3-4 dan computer opponent at your beck and call (on a home machine) you'd need a much more powerful home machine than most of us have. However I think that's not out of the question these days since this is pure crunch power, not total resources of a server class machine. Say something with an i7-2600, i7-3770, or better in it. I believe those would put you in the same ball park as the machines some of the stronger bots are running on instead of an order of magnitude weaker (while powerful server class machines and cluster probably older cpus that have been outclassed by newer ones).

2nd of all, at your current strength, even a 1 dan MCTS program should be strong enough and all of the stronger MCTS programs can deliver that on what would have been a strong home machine a decade ago (say a 2 core , 2 GHz machine) and many of us have machines a couple times more powerful than that because more modern CPUs.

Remember, you don't have to play taking a nine stone handicap. Even if this opponent at your beck and call isn't strong enough to force you to take more than 5-6 stones that's still "high handicap". It's not till you get down to 3 stones or less that the opening is "open".

I don't think you'd need more than a "decent" home machine till you got better than 1-2 kyu.