Computer Chess Club Archives


Search

Terms

Messages

Subject: Search Extensions When Probing Tablebases

Author: Ed Trice

Date: 22:56:42 11/18/05


I am wondering if any programs incorporate the concept of fractional ply search
extensions as leaf nodes are encountered that are "near" the tablebases. For
example, in my own program, Gothic Vortex, I have the complete 5-piece tablebase
available for probing, and about 128 MB of buffers set aside for this (overkill
is my middle name.)

As an experiment, I sampled some random positions encountered during the search
that had 6 and 7 pieces of the form:

1. Piece + Pawn vs. Same Piece + Pawn (+ kings of course)
2. Piece + Pawn vs. Same Piece + 2 Pawns
3. Piece + Pawn vs. Same Piece + Minor
4. Piece + Minor vs. Same Piece + Same Minor + 1 Pawn

These leaf nodes (6 or 7 pieces only) were hit from a distance, with about
12-13 pieces on the board total (all counts including kings.)

I had the program print out the position, and the score, for about 1000 random
positions in each of the 4 categories shown above.

I was amazed how often it would show a score of about +/- 1.3 pawns when one
side was literally just a few plies from swapping into the tablebases. Even more
shocking, in the 6-piece positions, the scores were +/- 0.2 pawns, and sometimes
dead wrong (showing close to a draw in R + P vs. R + P when one side could win
decisively!)

But, since these were leaf nodes, the search was stopped, and the less accurate
score was returned, and sent back down the tree.

In one extreme case, a score of -1.28 was returned as a leaf node is a Pice +
Pawn vs. Piece position, but when I set up that position and did a search, it
announced mate in 61 in a fraction of a second (because it searched and hit the
tablebases.)

So, I have been instructing the program to extend the search in certain leaf
nodes with 6 or 7 pieces, for the purpose of finding ways to hit the tablebases.
The results so far are mixed: much more accurate principal variations are
displayed in the endgames "near" the tablebases, but the tree tends to bloat by
a factor of 3 (too many extensions are not a positive payoff). Even with this
bloating, the time-to-depth, on average, is a 30% improvement over the
non-extended searches the closer you get to the tablebases (8- and 9-piece
positions, for example, are much better overall.)

I was wondering if any other programs employ search extensions when in close
proximity to tablebases?

My tablebases page:

http://www.gothicchess.com/javascript_endings.html



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.