Re: [computer-go] SlugGo approach: GNU vs.Goliath

["Jay Scott" <jay@xxxxxxxxxxxx>]

David G Doshay <ddoshay@xxxxxxxxxxxxxxxxx>:
>disclaimer warning: ALL OF IT IS PURE SPECULATION!

I have a different speculative theory, based on the nature of playout
search. It's a pretty vague speculation--not all the terms I use have
clear definitions--but it makes sense to me.

You can say that search works because it brings information from the
future back to the present, where it can be used to make decisions. A
wide shallow search brings back a lot of information about the near
future. A deep narrow search brings back a small selection of information
about the distant future.

Most of the reasons that people (including me) have given to suppose that
playout search should not work boil down to this: The selection of which
information to bring back from the distant future is expected to be poor.
There are many decisions to make about what paths to search, and no way
to be sure that the decisions are correct.

Here's my speculation: Gnugo's evaluator is much stronger in short-term
understanding than in long-term understanding. *Absolute* correctness
does not matter, *relative* correctness of short-term versus long-term
evaluation is what matters. In other words, I'm guessing that gnugo has a
fair idea of what's important immediately, and comparatively little idea
of what will happen in the game situation over time; it takes into
account what it sees in front of it, and lacks skill to usefully see
beyond the immediate future.

If this guess is right, then the playout search is leveraging the
evaluator's relatively good short-term insight to fill in missing
information about the more distant future. The short-term insight does
not have to be accurate enough to produce 16 good moves, it has to be
accurate enough to produce more information about the situation 16 moves
ahead than it destroys. (A bad decision taken along a search path can be
seen as destroying the information gathered down that path. A playout
search, one move wide, provides the worst case of that; a full-width
search does not make decisions and therefore does not lose information by
making bad ones--it just notices, oh, all the other moves were better.)

This theory predicts that search depth is key and branching factor is
less important. The playout search works because it is making available
information about the far future--perhaps poor information, but better
than was available before.

  Jay

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