Author: Dann Corbit
Date: 20:55:21 12/17/99
Positions where the computer changes its mind a lot, I call "sparklers."
Some programs, for instance, may change their mind 3 or 4 times on a single ply,
or 3 times on ply 7, 3 times on ply 8, four times on ply 9, etc...
It appears to me (not having made measurements but only from observation) that
*most* of the time, it is changing its mind back and forth between a very small
number of candidates. Consider this position:
rr4k1/3nppbp/b2p1np1/q1pP4/8/P1N2NP1/1P2PPBP/R1BQ1RK1 b - -
Crafty starts off liking Qc7, then abandons that move. From there forward, the
eval wobbles between Ne8 and Nh5:
8 3.68 -1.39 1. ... Qc7 2. Qc2 Nb6 3. Rd1 Bc4 4.
e4 e6 5. dxe6 Bxe6
8-> 8.24 -1.39 1. ... Qc7 2. Qc2 Nb6 3. Rd1 Bc4 4.
e4 e6 5. dxe6 Bxe6
9 11.26 -1.38 1. ... Qc7 2. Qc2 Nb6 3. Rd1 Nc4 4.
e4 Qb7 5. Na4 Qc7
9 16.42 -1.27 1. ... Ne8 2. Qc2 Bxc3 3. bxc3 Bc4
4. Nd2 Bxe2 5. Re1 Qb5 6. a4
9 23.01 -1.22 1. ... Nh5 2. Qc2 Bxc3 3. bxc3 Bc4
4. Nd2 Bxe2 5. Re1 Qb5 6. a4
9-> 23.32 -1.22 1. ... Nh5 2. Qc2 Bxc3 3. bxc3 Bc4
4. Nd2 Bxe2 5. Re1 Qb5 6. a4
10 34.48 -1.53 1. ... Nh5 2. Qc2 Nb6 3. Nd2 Nc4 4.
g4 Nf6 5. g5 Ng4 6. Nde4
10 39.68 ++ 1. ... Ne8!!
10 45.45 -1.04 1. ... Ne8 2. Qc2 Bxc3 3. bxc3 Nb6
4. Ng5 Na4 5. c4 Qc3 6. Qxc3 Nxc3 7.
Bd2 Nxe2+
10-> 49.38 -1.04 1. ... Ne8 2. Qc2 Bxc3 3. bxc3 Nb6
4. Ng5 Na4 5. c4 Qc3 6. Qxc3 Nxc3 7.
Bd2 Nxe2+
11 1:04 -1.24 1. ... Ne8 2. Qc2 Bxc3 3. bxc3 Nb6
4. Nd2 Qa4 5. Qd1 Bb5 6. Qxa4 Rxa4
7. Re1 e6 8. dxe6 fxe6
11-> 1:30 -1.24 1. ... Ne8 2. Qc2 Bxc3 3. bxc3 Nb6
4. Nd2 Qa4 5. Qd1 Bb5 6. Qxa4 Rxa4
7. Re1 e6 8. dxe6 fxe6
12 2:01 ++ 1. ... Ne8!!
12 2:27 -0.84 1. ... Ne8 2. Qc2 Bxc3 3. bxc3 Nb6
4. Bd2 Bxe2 5. Rfe1 Bc4 6. Rxe7 Nxd5
7. Rd7 Qa4
12 3:58 -0.81 1. ... Nh5 2. Qc2 Bxc3 3. bxc3 Nb6
4. Bd2 Bxe2 5. Rfe1 Bxf3 6. Bxf3 Nc4
7. Bc1 Qc7 8. Bg2
12-> 4:23 -0.81 1. ... Nh5 2. Qc2 Bxc3 3. bxc3 Nb6
4. Bd2 Bxe2 5. Rfe1 Bxf3 6. Bxf3 Nc4
7. Bc1 Qc7 8. Bg2
13 6:48 -0.87 1. ... Nh5 2. Qc2 Bxc3 3. bxc3 Nhf6
4. Bg5 Nxd5 5. Rfc1 N5f6 6. e3 Qc7
7. Rcb1 e6 8. a4
I know that there have been studies to show that chess programs continue to
change their minds as you go deeper and deeper, at about the same probability
per forward ply. What I am wondering is, "Has anyone done a study to find out
how frequently the new choice is one that has already been considered?" and
(perhaps more importantly) is there any indicator of stability where the choice
oscillates between a minority of the possibilities?
In particular, I am wondering if there is any computable curve of probability
for narrowing the candidates. I think (perhaps) this could be used for special
extension searching. Kind of like null move, but instead after we pass some
statistical test, we consider only 5 moves. Then after another test is passed,
we study only 3 moves, finally, for the last ply or so we might look at only two
candidates.
Notions? Arguments? Shrieking epithets?
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.