Hi,

My program tracks commonly occurring patterns and counts how many times they 
have occurred, counting each board in the game as one occurrence.
It then counts how many times the next play is in each location within the 
pattern, resulting in a percentage chance for each location. These percentages 
do not add up to 100% as the next play is not necessarily within the pattern at 
all.
The program then examines a board during play and for each possible play 
averages out these percentages, which does generally result in pointless moves 
such as first and second line in an empty area getting less than 1%, and some 
urgent moves more than 10%.
My program isn't Monte-Carlo based, but I do use the results to do machine 
learning based move suggestion.

Timothy Maguire

>Hi, 
> 
>I'm writing my phd thesis on Go and currently I'm on a chapter 
>"Adaptive playouts". 
>5 years ago I made an experiment: 
> - track average result of the playout if the pattern was present and 
>the move in the middle was played. 
> - use this statistics to guide further playouts 
>It was complete failure. 
> 
>Did any of you tried to do the same? 
>Or use such statistics to guide the tree? 
>Or anything remotely similar? 
> 
>I'm primarly interested in failed experiments, so don't be shy :) 
>You can e-mail me on priv if you prefer. 
> 
>Thanks 
>?ukasz 
>PS Maybe there was there a thread calling this technique CRAVE. 
 

_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to