[computer-go] IEEE Spectrum article by Deep Blue creator

[Ian Osgood]

On Oct 2, 2007, at 10:33 AM, Phil G wrote:
>
> On a slight different topic, for those of you with experience  
> writing an evaluation function for an alpha-beta search, do you use  
> the number of total moves played to weight different parts of the  
> evaluation function? For example, it is easier towards the end of  
> the game to know the certainly of certain points and use that as  
> the evaluation function. But at the beginning of the game, an  
> estimate of territory seems to be a better function (in light of  
> not knowing the certain of any or few points). How do you merge  
> these functions as the game transitions between middle game to end  
> game? (and as different parts of the board are in various stages too).
>
> Phil

This problem crops up in computer chess. The middlegame evaluation  
differs markedly from endgame evaluation. For example, in the former  
you wish to keep the king protected whereas in the endgame it becomes  
a mobile, active piece.  One technique is to perform both  
evaluations, then combine them continuously weighted on how much the  
current position is "middle-game-like" or "endgame-like". For  
example, it could be a linear combination dependent on the amount of  
material left on the board. In Go, it could be some other measure of  
the progress of the game, such as stone density or move number. It is  
important that the evaluations be combined continuously instead of  
via a sudden cutoff or threshold, in order to avoid border effects.

Ian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://computer-go.org/pipermail/computer-go/attachments/20071002/da2e73b8/attachment.htm