[computer-go] Making Java much faster

[David Doshay]

On 1, Dec 2006, at 6:15 AM, Wodzu wrote:

>
> ----- Original Message ----- From: "David Doshay" <ddoshay at mac.com>
>
>> Also, my data shows that if I doubled the time allowed for  
>> playing,  thus
>> "using" the time gained from faster execution for doing deeper   
>> lookahead,
>> the results did not improve, but actually got worse.

> I've also noticed that when i run my engine to play itself on  
> different
> depth level.
> The "deeper" player sometimes was defeated by his opponent with  
> lesser depth
> analysis. I guess he wasn't able to fidna good move because of the  
> horizon effect.

In self-play, I find deeper always has a greater win rate. I only  
look at
statistics over many games. Of course, individual games go both ways.

> Although I think when you run your engine with some more depth it  
> will vipe
> out its opponent from board;)

Nope, the effect is clear. If the lookahead is too deep against a  
different
engine (Many Faces of Go) then there is a lower win rate than a less
deep lookahead. It just does not help to base evaluation upon boards
that never occur.

I have checked lookaheads of 0, 2, 4, 6, 8, and 16 against GNU Go (same
engine and thus a very high hit rate on the lookaheads), and 0, 2, 3,  
4, 6, 8,
and 16 against Many Faces, where the game usually branches from the  
paths
that SlugGo checks by the 4th move.


Cheers,
David