Re: [computer-go] Evaluation functions and turbulence

[Erik van der Werf <E.vanderWerf@xxxxxxxxxxxxx>]

David Elsdon wrote:

> I was just thinking about quiescence and turbulence in game positions. 
> I take turbulence to be the opposite of quiescence. I think that 
> turbulence is intended to be a measure of how badly an evaluation 
> function will work in a particular position. Low turbulence means we 
> will get a reasonable evaluation, high turbulence means we will get a 
> poor evaluation and it may be a good idea to do more look-ahead.
>
> In go turbulence would be related to unsettled groups (among other 
> things), In many cases you would have an idea of the ways in which the 
> evaluation could be wrong. For example, if an unsettled group would 
> yield 4 points of territory if it lived and cost 30 points (in 
> prisoners and territory) if it died, and the evaluation was 7 ignoring 
> that group, then the actual evaluation would be between 3 and 37. This 
> leads me to the idea of evaluation intervals. This could lead to some 
> interesting ways to search the game tree.
>
> For example, if the evaluation interval of best move at ply 1 did not 
> intersect with any other evaluations then further search would be 
> unnecessary. If the evaluation intervals of other moves intersected 
> with the best move's interval then these should be searched more deeply.
>
> Any mileage in any of this ??

Sure, if such knowledge is reliable you can use it to narrow the 
alpha-beta window, pruning irrelevant parts of the tree. It is quite 
similar to what I did with the (internal) unconditional bounds in Migos. 
See 
http://www.cs.unimaas.nl/~vanderwerf/pubdown/solving_go_on_small_boards.ps.gz

Erik

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/