Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Hello from Edmonton (and on Temporal Differences)

Author: Jay Scott

Date: 13:46:19 08/01/02

Go up one level in this thread


On July 31, 2002 at 21:33:29, James Swafford wrote:

>One of my biggest concerns was the time to train a complex evaluator.
>I spoke with Rich Sutton about this today, and he convinced me that it's
>doable.

It could take tons of CPU time, but I don't think that's the biggest obstacle.
The hard part is not getting any one thing to work, it's getting all the myriad
things up to par. A performance program has many moving parts, and every part
must work smoothly; just one crew member who's uncoordinated will tip the canoe.

If you do find yourself short of cycles, think about ways to coarsely
parallelize your computation. If you find a good one, look for volunteers.
SETI@Home taught us that there are a lot of CPU hours out there for the asking.

Here's another speedup hint: TDLeaf learns from the principal variation, which
means it is using more information than an algorithm which learns only from the
root scores or the game result. In theory, you ought to be able to learn from
the whole search tree. Because of cutoffs, most nodes have very little
information to provide--but there are a lot of nodes in the tree. "All" you have
to do is understand what they tell you.

Good luck and work hard!



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.