Author: Robert Hyatt
Date: 13:42:08 01/20/99
Go up one level in this thread
On January 20, 1999 at 01:29:50, Jouni Uski wrote: >On January 19, 1999 at 16:13:53, Robert Hyatt wrote: > >>On January 19, 1999 at 01:08:14, Jouni Uski wrote: >> >>>On January 18, 1999 at 07:03:44, Amir Ban wrote: >>> >>>>On January 18, 1999 at 05:42:13, Jouni Uski wrote: >>>> >>>>>In my 47 position endgame test suite I have this kind of results in P90 10MB >>>>>hash 10 min per position: >>>>> >>>>>Fritz5 41 (!) no TBs >>>>>Mchess 7 39 no TBs >>>>>Fritz 4 37 >>>>>Nimzo 98 34 >>>>>Crafty 16.2 33 >>>>>Rebel 10 32 >>>>>Junior5 29 (!) >>>>> OK... I looked at this suite and ran it on my quad xeon. One minute per move, and get 39 right, whether I use tablebases or not. (the 3-4 piece positions are so simple they can be solved with a normal search, so the tablebases speed things up, but it is a difference between 1 second and .01 seconds in those cases. I looked at the 8 I got wrong at 1 min, and I am not sure that all of those are 'winning moves'. IE what I would like to see is (a) how move X wins and (b) how all the other moves don't win. In one case the key move was at ply=3 in the PV as it found another move it thought should be played first, and the score was still 'winning.' But as I looked at other results it became apparent that this is not only a tactical suite in many cases, but some positions are probably _not_ tactical. And then it is a question of positional judgement. Several I got right instantly based purely on positional evaluation. Some it _only_ got right on positional grounds with no tactical win in sight. Nothing wrong with a position, _if_ there is a reason for a move. IE in fine 70, white plays Kb1 or it draws. You can search to see Kb1 wins a pawn, or you can search to see Kb1 is the only move that keeps you on the coordinated square with your opponent. But clearly Kb1 wins, everything else draws. Some of these positions don't leave me with that 'feeling'... And I don't like tests that produce results like that unless it is to 'find this plan' where you have 2-3 moves that must be played, and so long as they show up in the PV, it is considered correct... Because you choose the right moves, whether for the right or wrong reason is a question of course... >>>>>Conclusion: Either 1) I have selected positions badly or 2) endgame skills >>>>>have not so much importance. But I have got from CCC feeling, that J5 is >>>>>good in actual endgame play. >>>> >>>>Can you repost the suite or give a link to it ? >>>> >>>>Amir >>> >>>5k2/8/8/8/8/2P5/8/3K4 w - - id "01 Kc2"; >>>8/7p/8/3K3P/8/8/8/5k2 b - - id "02 Ke2"; >>>8/5p2/2k5/K7/8/1P6/8/8 b - - id "03 Kd5"; >>>8/8/4K3/4P2p/8/5k2/8/8 w - - id "04 Kf5"; >>>5k2/8/3K1Pp1/6P1/8/8/8/8 w - - id "05 f7"; >>>5k2/8/8/1p5p/1P5P/8/4K3/8 w - - id "06 Kd3"; >>>8/7p/p7/1p1k4/8/3K4/1P4PP/8 w - - id "07 b4"; >>>8/8/1p1k4/5ppp/PPK1p3/6P1/5PP1/8 b - - id "08 f4"; >>>6k1/6p1/6Pp/ppp5/3pn2P/1P3K2/1PP2P2/3N4 b - - id "09 Nc3"; >>>8/5k1p/5P1P/1p6/bKpB4/P1P5/1P6/8 w - - id "10 b3"; >>>6k1/2b2p1p/ppP3p1/4p3/PP1B4/5PP1/7P/7K w - - id "11 Bxb6"; >>>8/8/8/5bk1/8/8/4B1PP/5K2 b - - id "12 Bh3"; >>>3b4/5kp1/1p1p1p1p/pP1PpP1P/P1P1P3/3KN3/8/8 w - - id "13 c5"; >>>8/8/8/8/8/8/pn2K3/B1k5 b - - id "14 Na4"; >>>5k2/1p6/p7/4pr2/8/P7/1P4P1/R5K1 w - - id "15 Rf1"; >>>8/8/2k5/8/4PK2/8/6R1/4r3 w - - id "16 Rd2"; >>>R7/4k3/8/8/8/p7/r3p3/4K3 b - - id "17 Ra1"; >>>5k2/p1p4R/1pr5/3p1pP1/P2P1P2/2P2K2/8/8 w - - id "18 Kg3"; >>>8/5k2/2p4R/5p2/8/P6P/1P4r1/1K6 b - - id "19 f4"; >>>6k1/1pR2p1p/1p6/3p1p2/4rP2/P3K3/1P4PP/8 w - - id "20 Kd3"; >>>2K5/2P2R2/k7/8/8/8/2r2p2/8 w - - id "21 Kb8"; >>>8/8/3q4/8/1P2K3/8/Q7/7k w - - id "22 Qd5"; >>>3K4/p2PQpkp/2q5/4P3/8/8/8/8 b - - id "23 Qb6+"; >>>k7/8/2p5/4P3/4B3/6p1/8/2n3K1 w - - id "24 Bd3"; >>>k1N5/2p5/8/1P6/8/8/7K/8 w - - id "25 Nb6+"; >>>2n5/kP6/8/K7/4B3/8/8/8 w - - id "26 bxc8N+"; >>>8/3Np3/7P/1p3P2/1k6/5K2/1b6/8 w - - id "27 f6"; >>>2r5/2pR1pk1/p1P3p1/P2K4/5P1p/5P1P/8/8 w - - id "28 Rxc7"; >>>8/1p5k/p1p4n/P6P/1P4N1/8/6K1/8 w - - id "29 Nxh6"; >>>8/6pk/1p6/8/PP3p1p/5P2/4KP1q/3Q4 w - - id "30 Qf1"; >>>8/2b2pp1/K2p4/p4P2/P7/6P1/8/2k3B1 w - - id "31 f6"; >>>8/4k3/8/6r1/1KP5/8/8/7R w - - id "32 Rd1"; >>>8/4kp2/4p1p1/2p1r3/PpP5/3R4/1P1K1PP1/8 w - - id "33 g4"; >>>3k4/3q4/8/4Q3/5P2/6K1/8/8 b - - id "34 Qe7"; >>>8/k7/8/2R5/8/4q3/8/4B2K w - - id "35 Bg3"; >>>8/8/2N4p/p5kP/P1K5/1P6/8/4b3 w - - id "36 Nxa5"; >>>8/8/pp6/3b2K1/1P6/PN2k3/8/8 w - - id "37 Na5"; >>>3R4/2N2p1p/5kp1/r7/pb6/4P1P1/5P1P/6K1 b - - id "38 a3"; >>>8/6Q1/p3r2p/3k3K/6P1/8/8/8 w - - id "39 Qxh6"; >>>8/7p/3k2p1/6P1/4KP2/8/7P/8 w - - id "40 h4"; >>>8/2p5/8/2kPKp1p/2p4P/2P5/3P4/8 w - - id "41 d6"; >>>7K/k7/8/P7/6pP/8/8/8 w - - id "42 a6"; >>>8/pKR5/8/p7/8/8/2p5/2k5 w - - id "43 Kb8"; >>>k7/7K/NR6/8/7p/6p1/p7/8 w - - id "44 Nb8"; >>>8/8/5p2/5p2/5P2/3p3B/5k1P/3K4 w - - id "45 Bg2"; >>>8/1p6/4p3/1K2k3/8/P7/8/8 w - - id "46 Kc5"; >>>4kr2/R4p2/6p1/8/1K6/3B4/8/8 w - - id "47 Bf5"; >>> >>>Note, that DarkThough can solve all 47 in 10 minute level! So they are probably >>>all correct. >> >> >>That's not a convincing argument. A computer can find bad moves just as >>efficiently as it can find good moves. I had hoped this was an endgame test. >>But it sounds more like a tactical test based on endgame positions. IE just >>another WAC test suite. I'd like to see more tests without a tactical solution, >>where you have to play a move because it is the _right_ move, not because it >>wins tactically. IE like the non-tactical half of the Bratko-Kopec test... > >But why is Junior so poor in this suite, when tactically it's normally very >strong and up to Fritz?! Good question. One is an underpromotion, so it won't get that one. :) But I'll bet that other programs are solving many of them purely on positional considerations and not tactical ones... which means 'judgement' is needed. And _if_ the key moves are not best tactically, can we be sure they are best positionally??? It might be that Junior is doing better than anyone here, because some of those moves might be bad...
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.