[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: A problem with understanding lookahead
I don't think the intuition you're appealing to in your "thought experiment"
really proves much.
There's a lot hidden in the phrase "reasonable" evaluation function. In
chess, counting material seemed reasonable, but it was far from clear that
this could ever lead to the sophisticated positional judgement that strong
players all believed was absolutely necessary for high level play.
Now take go. Let's consider a reasonable (but naive) evaluation function:
count territory, assuming everything on the board is alive. It's not a
perverse evaluation function; on the contrary if you could search all the
way to the leaves, it's actually "correct." But it's far from clear that,
say, a depth 6 full-width search using this evaluation function will yield
even a 20 kyu program, or that searching to depth 7 (or 8 or 9) will be
significantly better than searching 6. How deep do you have to go before the
program can appreciate the value of a crosscut that doesn't lead to an
immediate capture? How deep do you have to search before it sees the value
of strengthening a weak group that's got lots of liberties but no eyes?
The fact that /we/ would like to be able to do full lookahead with /our/
evaluation functions doesn't mean that any player with a "reasonable"
evaluation function will benefit from limited search, even in games we think
of as having "well behaved" game trees.
-David Mechner
----- Original Message -----
From: "Don Dailey" <drd@xxxxxxxxxxxxxxxxx>
To: <computer-go@xxxxxxxxxxxxxxxxx>
Sent: Wednesday, January 17, 2001 6:08 PM
Subject: Re: computer-go: A problem with understanding lookahead
>
> Yes, if your evaluation is totally out of whack I agree with you. The
> power of the search is that it helps you reach your goals, even if
> your goals are stupid and out of line. You might regret reaching your
> goals, but you get better at it with increasing search depth.
>
> So I think the thought experiment is valid. If your evaluation is
> this far out of whack, then your program really just doesn't have a
> clue about what is good and what isn't and thus a 1 ply search does
> the least damage. In fact, maybe you should just change the sign of
> the evaluation!
>
> Don