[computer-go] Computer go in 7x7

[Rémi Coulom]

sylvain.gelly at m4x.org wrote:

>Hello all,
>
>I am wondering about Go in 7x7. I know that the game in this size has no real 
>interest in itself. However, I think that the level of computer go programs 
>is much higher in 7x7 than in 9x9, and it could be interesting to see until 
>where we can go in 7x7.
>
>There was a discussion on this list recently about the level of CrazyStone in 
>7x7, and also (perhaps not on this list, only on kgs I don't remember) of 
>Magnus saying that the level of Valkyria was very good on 7x7. I am wondering 
>if we can make a player that beats almost all humans in 7x7?
>  
>
I think Crazy Stone does, already. Look at its history on KGS. It won 
against strong dan players, and even a 5-dan professional. Crazy Stone 
works with wine in Linux, if you wish to try it by yourself.

>I have tried to put MoGo on kgs in 7x7, but it looses quite a lot of games. 
>Ok, I have put only 5 minutes time, so it plays quite fast, and perhaps could 
>be better with more time, but it is clearly far from a really good player.
>  
>
A lot of the strength of Crazy Stone comes from its hand-made opening book.

>I have implemented a meta-UCT (that means that instead of playing a random 
>game after the tree, you make MoGo playing against itself) to generate an 
>opening database. I have made this meta-UCT 100% parallel for a cluster (as 
>the evaluation is now so costly, parallelisation is trivial). I can dedicate 
>quite a lot of computers to this task as I have access to a cluster.
>So I wonder if you think that this meta-UCT can be effective to generate a 
>good opening database, and this way manage to have a really good computer Go 
>player in 7x7 ? 
>Perhaps usual UCT exploits too much for this task, because my experiments show 
>that the beginning of the tree is quite narrow. Here are the first moves: D4 
>D5 E5 C4 D3 C5. I have no idea if it is very stupid or normal.
>
>Remi was saying that the right komi for 7x7 is 9, but I have made my 
>experiments with komi 7.5, and the meta-UCT predicts that black is almost 
>always loosing. Is it because the blacks moves are very bad?
>  
>
Yes. I am very confident that 7.5 is too small.

By the way, using an integer komi is interesting because the possible 
outcomes are not binary any more. The game can be a win, a draw, or a 
loss. I have the feeling that UCT works really very well when the 
outcome is binary. Introducing the possibility of a draw significantly 
changes the nature of the problem. My feeling is that UCT becomes too 
selective in that case.

>To conclude, I think this question of 7x7 is interesting because we can 
>perhaps have a very good player. What do you think? If you have good ideas to 
>make the meta-UCT generate a good database, I can dedicate a lot of CPU to 
>this task, and would interested to see what happen.
>  
>
I don't believe it can beat hand-made book preparation, but I may be wrong.

Rémi