I have a few people I would consider to be my rival in the world of go, but there is only one computer AI... Fuego.

It's style is... unusual, but it has been effective at beating me. In our dozens of games, I have beaten it three times. Once when I allowed myself to undo some bad moves, once when I had a 5 stone handicap (but no undos), and once by playing mirror go as white (it just kept playing normally so it resigned because I was going to win by komi).

Now that I'm getting better, I want to beat Fuego fair and square. No hanicap, no undo, and no tricks.

Any help you guys can give me would be appreciated. I would ask that if you know any quirks with the AI (like moves that it always gets wrong) that you wouldn't post them here. I want to beat Fuego by improving my game, not by learning weaknesses that only it has.

Tonight, I played the first in a series of games I intend to do against this AI with the hope of winning an even game. I included a few comments, but mostly this is just the game. Fuego beat me again. I was wondering if you guys could tell me what I'm doing wrong when responding to its unorthodox moves.



For those wondering, this AI is open source and totally free, you can get it here - http://sourceforge.net/projects/fuego/ It is also available to play against on the KGS as fuego19 (where it is currently rated as 1 dan).

EDIT: Also, the hardware this computer is a Core i5 2.5GHz with 6 GB DDR3 RAM.

_________________
"You have to walk before you can run. Black 1 was a walking move.
I blushed inwardly to recall the ignorant thoughts that had gone through
my mind before, when I had not realized the true worth of Black 1."
-Kageyama Toshiro on proper moves

I don't think that's how MC bots work. Do you know the basic MC algorithm?
The only weakness of MC bots seems to be wasting ko threats for no reason what's-o-ever, but I don't think you can use that as a strategy against such bots.



Me, too. I play it on 9x9 only. On such a small board and my hardware it's somewhere between kgs 3d-5d and I haven't even come close to beating it in an even game. I usually take black without komi.


On bigger boards it's obviously much much weaker. On 19x19 you just have to keep an eye on the center and make sure your groups don't die, pretty much.

But it's also a question of reading power, which scales proportionally with the processing power/memory you give Fuego and reciprocally with boardsize.

On 9x9 with strong hardware Fuego has even beaten some pro (edit: Zhou Junxun 9d from Taiwan, see senseis), so not sure there's a "secret" to defeating it, other than just getting stronger in general.

Strange fuseki moves are another one. Once you get Fuego out of its "opening book" it can play some odd moves on the 7th/8th line, tenuki for little profit from an approach to a hoshi stone and similar.

Edit:

For example a game where I was messing around today: (i5 2500k 3.3 GHz, 8GB for reference)



and all look strange to me. Up to was a (common) pro sequence though.

For a while when I was getting started with go, Aya 6.34 was my rival. It played a bit more normally than fuego.

I guess your plan should be pretty clear.

Thanks for sharing - I didn't know that it came with a gui. Nice. I played it once and it made some moves that even I could see were awful - but from what I've heard, mc engines can learn from their mistakes. This one however has a lot to learn 'till it gets to 1d. Here's the game:




Here are a few comments. Take with a grain of salt, I'm not much stronger than you.

_________________
Patience, grasshopper.

MC bots like Fuego are quite sensitive to what hardware they're running on, moreso than traditional bots. Do you know the processor and memory of the machine you played it on Daal?

We can be more specific.

a) AFAIK none of the MC based programs are even 1 dan on a machine with less than a dual core 2 GHz processor. The same program that can play at a few dan on a powerful workstation/server class machine (say an I7-2600, i7-3770, or better CPU in it)isn't going to be even 1 dan on a tablet.

b) The weaker settings of such programs might not be using the MC algorithm. It simply isn't possible to get this algorithm to be other than erratic if working at worse than a couple kyu.

The big thing I noticed in that game is that you had a lot of opportunities to reduce the center, but tended to get too attached to taking a few points on the side, preventing reductions of only about 10-15 points. Meanwhile, any invasion of the center was worth at least 30 points.

Center territory is vulnerable, but if you don't make the moves to destroy it, it's also gigantic.

_________________
Tactics yes, Tact no...

Thanks, daal. You are right, I should have played on the right before moving on the bottom. I was forgetting what I learned in The Direction of Play. Next time I'll be more conscious of my opening. Congratulations on beating Fuego.


Interesting. It is true that I am fond of side-oriented strategies. One of the reasons I play the low Chinese so much is that it often turns into a big side for me. (That or a fun game of running my opponent's weak invasion group around the board) When I first started playing I was very territorial, preferring a san-san, komoku fuseki that secured the corners and built along the side before challenging the center - a part of me still likes that style of play.

Perhaps Fuego can teach me the value of the center. I was taught the traditional approach of corners, sides, then center, so I will naturally play that way. I was thrilled when Fuego was giving me 4th and 5th line territory on the top because I know that is huge, but I wasn't paying attention to the fact that he was giving it to me in exchange for a massive wall defending the center. Traditional wisdom says that pushing your opponent to make 4th and 5th line territory is bad because for every stone they get 3-4 solid points. However, Fuego knew that his center was worth more and didn't care about proverbs or "traditional wisdom."

Part of me says I should channel my inner san-ren-sei and forget everything I hate about that opening, but I will not. But that's not saying I won't adapt my go for the next game...

_________________
"You have to walk before you can run. Black 1 was a walking move.
I blushed inwardly to recall the ignorant thoughts that had gone through
my mind before, when I had not realized the true worth of Black 1."
-Kageyama Toshiro on proper moves

i5 2,40 MHz, 6 GB ram.

_________________
Patience, grasshopper.

Fuego is pretty strong, sort of. I played it on an i7-3520M laptop (dual core 2.90 Ghz machine with 8gigs of RAM). After 100 moves it was well ahead, in my opinion, although I wasn't playing too hard (I would say that of course ), but then, instead of consolidating it's lead, it rather collapsed, and things just went from bad to worse, as is the norm with MC. It eventually resigned when I was about 60 points ahead.

Takemiya is also known for playing moves that are conventionally thought of as bad in this way, but his style is quite hard to play. The reason it's "bad" is that it assumes that an opponent will play to limit the center at some point so that it's hard for the center-oriented side to make as many points there as they give away, but if it's never challenged, that isn't an issue

I'm not sure if this helps, but when chess programs were still beatable the way to do it was always almost a closed position.

The go equivalent to this would probably be to stay away from fights where the one wins that can read further/that come down simply to who can calculate better, since there even home pcs are miles better than any human will ever be.

Not 100% sure though, take it with a grain of salt =)

Go is still very young in AI terms. On a home pc you're talking mid amateur dan at absolute best, on server gear perhaps 6d amateur? Maybe not even that.

Sorry, I misread your post. I'm very tired.

I think actually monte carlo based bots often struggle in cases where capturing races are involved,perhaps just the cases when there are multiple capturing races on the same board. The bot is unable to compartmentalize in the same way that a human can. Things that are easyish for humans, like recognizing "this series of moves is forcing, and will be whenever it is played" are hard for bots in some cases.

_________________
Occupy Babel!

The Fuego-Moyoaji Ladder Game

I haven't come back to this thread in a while, but I have still been playing Fuego every once in a while. I still haven't won an even game, but I am getting closer.

The other day, Fuego played a very bizarre game with me - it played out a broken ladder in exchange for the weak influence it gives.

The whole reason I felt comfortable playing 55 was because I knew the ladder worked for me. The whole way down the ladder I kept reading it out and looking for where I was messing up. I must have spent 5 minutes just in that part of the game reading and re-reading to make sure I wasn't misunderstanding how the ladder worked for me. Fuego wouldn't play a bad ladder, right? At the end, the ladder was broken, but I was so taken aback that I didn't know what to do. I was so frustrated and confused that I actually ended up losing the game trying to find a way to take advantage of white's supposed mistake. Psychological warfare at its best. At move 245 I made a tragic self-atari and I gave up. I had to go to class anyway. So I took a break for a couple of hours then, after class, I tried again. I undid that blunder and then a few moves here and there toward the end of the game but was still unsuccessful.

Finally, I went all the way back to the original failed ladder. I knew I could win from there. I took a good long look and figured out a winning strategy from that point on. At move 187 Fuego resigned. Apart from a misclick at 157 (I was playing that one on a bus) I only needed that one undo at 87 to win. The game was still bizarre to say the least...

_________________
"You have to walk before you can run. Black 1 was a walking move.
I blushed inwardly to recall the ignorant thoughts that had gone through
my mind before, when I had not realized the true worth of Black 1."
-Kageyama Toshiro on proper moves