Re: [computer-go] Learning : was Chess programs versus go programs

[Chris Fant <chrisfant@xxxxxxxxx>]

> The real debate should be
> whether one can make a good enough evaluation function without using local
> tactical search inside the evaluation.  That's what makes evaluation so
> slow.  I believe local search is required since if tactical stability of
> stones is not established, there is no way to estimate territory.

This is how I started (no tactics), just to see how far I could take
it.  I still haven't implemented tactics, but I almost started it
today.  Clearly there is huge gains to be had just by doing the
smallest tactical searches.  That is one of the optimization problems
to solve: how hard to try for tactical understaning of each group
versus how many full-board nodes per second.

As for my soon-to-be-implemented tactics search, I know that I will
have to implement both A/B and PN so that I can compare them in the
context of my learning system.  I will do A/B first since I understand
it better (never tried PN for anything).  I expect to have good
ordering as well as information on how close a node is to the goal, so
it seems either could work well.  I always put off big steps like
this, but then once I get involved it goes much faster than expected.

I haven't thought about PN search management much yet.  Any words of
wisdom from the group?  Tree representation thoughts?.
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/