Author: GuyHaworth
Date: 16:24:56 02/06/06
Go up one level in this thread
Certainly, there are some situations where the DTM(ate) EGTs would give less
than optimal advice:
1) The EGTs are for a variant of Chess, namely (Chess minus Castling). Chess
studies have been devised where the stipulated result can only be achieved by
castling.
2) DTM EGTs do not recognise the 50-move rule, which (it is always worth
remembering) does not dictate that the result is a draw after 100 plies without
capture or P-push but merely entitles the defender to claim a draw in those
circumstances if they wish to.
DTZ50 EGTs do assume the defender does so, and will flag positions as draws if
capture/P-push can be delayed for more than 50 moves assuming minimaxing play.
Unfortunately, this does mean that (e.g.) btm KNNKP positions requiring 101
plies for Black to be forced to push the Pawn (i.e. 51 loser moves but only 50
winners moves) will still be considered wins in the DTZ50 EGT. There's a price
to pay for measuring depth in winner's moves rather than in plies and that's it
:-)
Two recent papers review the computation of non-DTM EGTs by John Tamplin,
latterly using Marc Bourzutschky's 'gtbgen' generalisation of Eugene Nalimov's
'tbgen' program:
Tamplin, J.A. and Haworth, G.McC. (2003). Chess Endgames: Data and Strategy.
Advances in Computer Games 10, Graz, Austria (eds. H.J.van den Herik, H. Iida
and E.A. Heinz), pp. 81-96. Kluwer Academic Publishers, Norwell, MA. ISBN
1-4020-7709-2.
Bourzutschky, M.S., Tamplin, J.A. and Haworth, G.McC. (2005). Chess endgames:
6-man data and strategy. Theoretical Computer Science, Vol. 349 Issue 2, pp.
140-157. ISSN 0304-3975.
DTC, DTZ and DTZ50 EGTs have been computed for all 3-to-5-man and 6-man P-less
endgames. How this data is used to filter the available moves to a smaller set
of 'optimal' moves is another question, also covered in the two articles above.
There are certainly 'pathological' positions where none of SC-, SM- and SZ-
(strategies respectively minimising DTC, DTM and DTZ) pick the move required to
retain a win in the context of a 50-move rule. In these positions, only SZ50-
does that, e.g.:
KNNKP: 8/8/8/8/2K3Np/7N/3k4/8 w ... 1. Ng1?? h3 and DTZ = 61m - draw
1. Ngf2 ... and White can still win
KQPKQ: 8/8/1P5Q/1K6/3q4/8/5k2/8 w ... 1. b7?? and DTZ = 51m - draw
1. Qg5! and later ... 50. b7 - win
KBBKNN: 8/8/6n1/8/k3BB2/8/n1K5/8 w ... 1. Bxg6?? and DTZ = 54m - draw
1. Bd6! and now DTZ50 = 27m - win
See the multi-metric site http://chess.jaet.org/endings/ for the detail.
A passing thought is that, if the attacker is playing SZ*Z50- (guarding the
length of the current phase but otherwise optimising DTZ50) and the defender is
playing SM+, less than optimal as pointed out, an EGT could be created to
optimise the winner's play using the 'SM+' knowledge - and no doubt, some
positions could be finessed as wins when they should only be draws. What would
we call such a database - an 'SZ*Z50-/SM+' EGT? Cumbersome but necesssary :-(
And SZ*Z50- is not the optimal strategy to play against a fallible player, but
that's another story (also told in the above articles).
As to 'time wasting', I don't see why chess engines cannot delegate their
EGT-enquiries to an asynchronous process on the side of the search, treating
returns from the EGTs as a bonus.
Thus, in principle, there is no reason why EGTs should subtract from the sum
total of silicon understanding.
And no game was drawn or won by resigning, so there's no reason for a
chess-engine to resign particularly, and less reason to blame it on EGT input.
g
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.