A match against Fuego

tsuhre · #1

It's been bothering me for quite some time that I can't seem to make any headway against the go playing program Fuego (an attachment of the gogui). I think even in the best match I even came close to winning legitimatley on ( by legitimately I mean never going back to fix mistakes) I still lost by 7.5 moku.

This is incredibly disheartening for me, especially since I've heard that intermediate level players can regularly beat the computer programs with ease. So, In a fit of frustration yesterday, I gave myself 9 handicap stones just to get some enjoyment out of finally crushing the computer. Unfortunately for me, Fuego (W) resigned after a blacks play at 123, so I still don't get any satisfaction form this match.

I'm posting it here, not because it was a particularly good or inventive match, but because I think I made relatively few mistakes throughout the course of the game. Let me know what you guys think.

billywoods · #2

I'm only 4 kyu, but a few notes:

Move 3: notice the strange shape that F8 and D10 make. It has a weak point in the middle. This isn't a problem right now, but you don't know what's going to happen in 100 moves' time. There's no point setting yourself up for trouble, and it's too early to say that this stone makes points or anything. There were a few shapes like this in the game.
Move 11: slow, endgame. Don't worry about white T3 or S2: they're big later on, but they don't make any eyes for white. Try pincering, e.g. at R7 or R8 - this will deprive white of a base and make the group squirm, and at the same time it develops the right-hand side for black.
Move 25: It's hard to see what you're doing here. You're not cutting the white stone off, or depriving it of a base, or making points, or helping your corner group. In fact, assume O4 and K8 both live, and imagine what this will look like in endgame; aren't you just playing on dame? There were a few points like this in the game.
Move 35: the tiger's mouth is a weak shape because of white's stone at O4 here (which can push in at O3 and cut you off).
Move 49: not urgent. Big, of course, if it keeps white's groups separated. (Though P5 can escape anyway even if black tenukis once.) More urgently, though, what's going on with your corner? (Tsumego time! Hint in the spoiler below...)
White R2...

Dusk Eagle · #3

Seeing how much you lose to a computer by is not a good metric. Most decent computers will only play so as to win by a small margin, and won't focus on increasing their lead when they're ahead like a typical human would. So you and I could both play to it and lose by the same amount of points. Also, computers have gotten quite a bit stronger over the past few years. The top computer program plays at a 6d level now, and Fuego plays at a 1d level if I'm not mistaken. They're definitely tougher to beat than before.

I wouldn't recommend using Fuego as your training bot. Instead, I think you should make an account on KGS (if you haven't) and play there. There are other beginners on the site, as well as weak bots that are almost always available for you to play against (a lot of bots on the site don't even play ranked games, so you don't have to worry about rank if that scares you). I think you'll learn a lot more from that (and from reading sensei's library, if you aren't) than from playing Fuego.

Anyway, billywoods covered what I wanted to cover regarding the actual game. I think it will be particularly helpful if you pay attention to his comments for move 11.

lovelove · #4

Your game is actually a 10 stone handicap game, and you have a bad habit of following your opponent's moves.

tsuhre · #5

billywoods wrote:

I'm only 4 kyu, but a few notes:

Move 3: notice the strange shape that F8 and D10 make. It has a weak point in the middle. This isn't a problem right now, but you don't know what's going to happen in 100 moves' time. There's no point setting yourself up for trouble, and it's too early to say that this stone makes points or anything. There were a few shapes like this in the game.
Move 11: slow, endgame. Don't worry about white T3 or S2: they're big later on, but they don't make any eyes for white. Try pincering, e.g. at R7 or R8 - this will deprive white of a base and make the group squirm, and at the same time it develops the right-hand side for black.
Move 25: It's hard to see what you're doing here. You're not cutting the white stone off, or depriving it of a base, or making points, or helping your corner group. In fact, assume O4 and K8 both live, and imagine what this will look like in endgame; aren't you just playing on dame? There were a few points like this in the game.
Move 35: the tiger's mouth is a weak shape because of white's stone at O4 here (which can push in at O3 and cut you off).
Move 49: not urgent. Big, of course, if it keeps white's groups separated. (Though P5 can escape anyway even if black tenukis once.) More urgently, though, what's going on with your corner? (Tsumego time! Hint in the spoiler below...)
White R2...

Thanks for the tips! This was exactly what I was looking for just in terms of amateur mistakes I was making on the board. I thought I might just run through some of what you talked about and what I was thinking at the time.

[list][*]Move 3: Ah yes, the elephants eye. I actually regretted this move as soon as I played it.
[*]Move 11: I realize now that I didn't actually have to play this move. (actually I realized this mid-game when the same situation occurred in the lower left corner. I'll admit that I never even considered R8 or R7 which seems silly in retrospect.
[*]Move 25: I have no idea what I was doing here either. To be honest it was not a particularly rational match.
[*]Move 35: I think I realized this and was more concerned with cutting white off. Looking back though, if that was my goal I should have just played Q5 directly
[*]Move 49: I think I was concerned with preserving shape here. Is a white play at R2 tsumego? I would probably respond with a hane at R1.

Dusk Eagle wrote:

Seeing how much you lose to a computer by is not a good metric. Most decent computers will only play so as to win by a small margin, and won't focus on increasing their lead when they're ahead like a typical human would. So you and I could both play to it and lose by the same amount of points. Also, computers have gotten quite a bit stronger over the past few years. The top computer program plays at a 6d level now, and Fuego plays at a 1d level if I'm not mistaken. They're definitely tougher to beat than before.

I wouldn't recommend using Fuego as your training bot. Instead, I think you should make an account on KGS (if you haven't) and play there. There are other beginners on the site, as well as weak bots that are almost always available for you to play against (a lot of bots on the site don't even play ranked games, so you don't have to worry about rank if that scares you). I think you'll learn a lot more from that (and from reading sensei's library, if you aren't) than from playing Fuego.

Anyway, billywoods covered what I wanted to cover regarding the actual game. I think it will be particularly helpful if you pay attention to his comments for move 11.

Thanks for the advice! I had suspected that fuego played pretty minimally for a while since the score was always so close. I completely agree though, It's a pointless exercise to compare yourself in skill level to a computer program. What I'm really concerned with here is how I approached individual situations, and what I could do to make those approaches stronger. (as you say, billywoods covers this excellently)

I'm suprised though: I'll admit that I never really thought playing endgame moves like 11 was that detrimental to the rest of the board.

I'm still working myself up to try out KGS. I've been playing on Flyordie for about 6 months now just because it was an easier GUI for me to work with.

lovelove wrote:

Your game is actually a 10 stone handicap game, and you have a bad habit of following your opponent's moves.

Hahaha, that makes these mistakes even worse then. In this game I almost exclusively followed my opponent, which I know is terrible form. In my defense, I would never play this way in a live match, but I would also never have 10 handicap stones in a live match.

topazg · #6

Some thoughts

tsuhre · #7

topazg wrote:

Some thoughts

Thanks a bunch topazg

I appreciate the input.

xed_over · #8

tsuhre wrote:

Unfortunately for me, Fuego (W) resigned after a blacks play at 123, so I still don't get any satisfaction form this match.

most programs have an option to "not resign". You may have to search for it.

tsuhre · #9

xed_over wrote:

tsuhre wrote:

Unfortunately for me, Fuego (W) resigned after a blacks play at 123, so I still don't get any satisfaction form this match.

most programs have an option to "not resign". You may have to search for it.

Good to know

Mike Novack · **#10**

Do not turn off "allowed to resign" when playing against any of the MCTS based programs.

Forcing the program to play on when it has concluded "hopeless" will not give you much satisfaction nor will it give you an idea of the margin of difference. Once there are no moves with probability of winning higher than other moves it will begin making truly silly moves.

Remember, it ignores "by how much" (won or lost) and so forced to play on it isn't making plays that will cause hold down the margin of defeat.

oren · **#11**

Mike Novack wrote:

Forcing the program to play on when it has concluded "hopeless" will not give you much satisfaction nor will it give you an idea of the margin of difference. Once there are no moves with probability of winning higher than other moves it will begin making truly silly moves.

They have programmed in VERY human behavior then.

xed_over · **#12**

Mike Novack wrote:

Forcing the program to play on when it has concluded "hopeless" will not give you much satisfaction

it gives me satisfaction

tsuhre · **#13**

In a way it's somewhat irrelevant anyway. I might eke some short term satisfaction out of crushing the computer, but it wont help my overall strength. As mike says, changing the option to do not resign changes how the computer plays and is therefore not a good determining factor in how much I've surpassed it.

billywoods · **#14**

tsuhre wrote:

Is a white play at R2 tsumego?

What I meant was that your corner group was very small, had few liberties, and had a cutting point - there's no guarantee it's alive, unless you've read it out to check that it's alive. If you're strong enough, you should do this sort of reading as a matter of practice (reading problems are called tsumego, and are often simple corner life-and-death problems anyway). If you're not strong enough, you should definitely at least notice that you're close to dying, and consider putting in an extra securing move. I've put some thoughts in the spoiler below to show you how I'd think it out, but if reading is still very difficult for you, don't take them too seriously.

(I think it's a very nice reading problem - basically every move is an application of the proverb "your opponent's best move is your best move". If you're not yet strong enough to read these things out for yourself, it may be worthwhile for you to spend 10 minutes running through my solution below (in your head, not on a board, for maximum benefit!).)

There are only really a few classes of move to consider. Supposing for a second that white plays tenuki, black R2 makes one eye instantly and threatens to make another eye on the next move. If white doesn't play tenuki, you can check yourself that if white plays any one of Q1, R1, R3, S1, S2, T1 or T2, then black R2 is a good response (that makes black live). (Be careful that neither your T3 stones nor your Q2 stones run out of liberties, otherwise you'll have to fill that eye back in. There's some weird-looking stuff with some of the moves above, but ultimately none of it is harmful.)

So, R2 is the move that black really wants, so it's the only interesting move for white (as it's the only one that doesn't fail outright so far!). White R2 is threatening to connect underneath on the next move (e.g. R2 S2 Q1 R3 R1 S1 P1, and black only has a single two-space eye; a throw-in doesn't help, because R2 S2 Q1 R1 S1 R3 Q1 T1 still gives white time to connect out - check that similar things happen if white plays R2 and black responds any of R3, S1, T1 or T2). So we now only have two responses to consider for black: Q1 and R1.

Let's suppose black responds R1. Intuitively, you probably agree that white's interesting moves are the hane (atari) and extending (to S2) - this is more or less because black would like to be able to play S2 on the next move, but these are the two moves that stop that being possible. If R2 R1 S1, then it's unreasonable to allow white to take the stone, so black connects at Q1. Now white can just connect, and the group is dead. (If you're not sure why: later white could atari at P1, black R3, white T2, black captures at T1, and white has a dead shape inside.) If R2 R1 S2, then black S1 allows white to make a dead shape with T2; alternatively, black T2 allows white to make a dead shape with S1. So black R1 - your suggested move - dies.

Let's suppose black responds Q1. If white plays any of R1, R3, S1, T1 or T2, then S2 is a nice eyeshapey move and black lives. So of course white wants to try playing S2. Now if black plays R1, R3, T1 or T2, then white S1 makes a dead shape; so black had better play S1. After this, black T2 would make an eye (check there are no liberty problems), so white plays T2 and prevents it; ~~the best result black can get here (and hence overall) is to connect at R3 for seki~~ (check what happens if black R1 or T1; white should then play T4). Edit: nope - no seki. See Dusk Eagle's post below. Oops!

So - provided white makes no mistakes - black ~~has exactly one way to live in seki, and everything else~~ dies outright. I think.

Edit: typos...

Dusk Eagle · **#15**

Black can't even make seki.

Click Here To Show Diagram Code: [go]$$Wc $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . X O O O . | $$ | . . . , . . . . . , . . . O O X X O a | $$ | . . . . . . . . . . . . . O X X 6 X X | $$ | . . . . . . . . . . . . . O O X 1 3 5 | $$ | . . . . . . . . . . . . . . b 2 c 4 . | $$ ---------------------------------------[/go]

This may look like a seki, but notice what happens after white plays 'a' and 'b'. Black suddenly has one liberty, so he dies. Of course, black could play 'c' at any point before white gets 'a' and 'b', but then black is left with a dead shape inside, so he's dead.

karaklis · **#16**

Dusk Eagle wrote:

Fuego plays at a 1d level if I'm not mistaken

Last time I downloaded it, it was around 6k. Has it improved so much since then?

billywoods · **#17**

Dusk Eagle wrote:

Black can't even make seki.

Click Here To Show Diagram Code: [go]$$Wc $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . X O O O . | $$ | . . . , . . . . . , . . . O O X X O a | $$ | . . . . . . . . . . . . . O X X 6 X X | $$ | . . . . . . . . . . . . . O O X 1 3 5 | $$ | . . . . . . . . . . . . . . b 2 c 4 . | $$ ---------------------------------------[/go]

This may look like a seki, but notice what happens after white plays 'a' and 'b'. Black suddenly has one liberty, so he dies. Of course, black could play 'c' at any point before white gets 'a' and 'b', but then black is left with a dead shape inside, so he's dead.

You're right. In my hurry to read it out and put a full solution together, I was too hasty. Thanks.

tsuhre · **#18**

Dusk Eagle wrote:

Black can't even make seki.

Click Here To Show Diagram Code: [go]$$Wc $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . X O O O . | $$ | . . . , . . . . . , . . . O O X X O a | $$ | . . . . . . . . . . . . . O X X 6 X X | $$ | . . . . . . . . . . . . . O O X 1 3 5 | $$ | . . . . . . . . . . . . . . b 2 c 4 . | $$ ---------------------------------------[/go]

This may look like a seki, but notice what happens after white plays 'a' and 'b'. Black suddenly has one liberty, so he dies. Of course, black could play 'c' at any point before white gets 'a' and 'b', but then black is left with a dead shape inside, so he's dead.

Thanks for the solution. It would make a good life or death problem.

hyperpape · **#19**

karaklis wrote:

Dusk Eagle wrote:

Fuego plays at a 1d level if I'm not mistaken

Last time I downloaded it, it was around 6k. Has it improved so much since then?

Last time I played, it was better than me. Time settings matter, of course.

Mike Novack · **#20**

I perhaps wasn't being clear enough by what I meant as "silly".

To put it into words, the MCTS algorithm results in "make the move that has a higher probability of winning than other moves". But if the probability of winning for all moves is close to zero then the differences between these probabilities are also close to zero and there is no way for the algorithm to distinguish between "hopeless, but not making things worse" and "hopeless and making things worse".

That is why important to always allow these programs to resign. It's not so much that the program has concluded "can't win the game" but also "I can't distinguish between sensible and ridiculous moves". In other words, the MCTS algorithm doesn't work in this region.

If you have experience playing against these programs you will note that a human might resign earlier. Toward the end of a game that it will lose the program will make "insulting" attacks, play moves that will work only if responded against very poorly or not at all. The reason is that from its calculation the probability of that isn't zero. With experience playing against these programs you will learn to recognize when close to the end of the game whether the program "thinks" it will win or lose. Unnecessary safety moves that cost a point or two means the program has the game in hand while attacks that would work only against a bad mistake means the program "knows" it has lost.

A match against Fuego

Who is online