Author: Robert Hyatt
Date: 07:27:20 01/24/01
Go up one level in this thread
On January 24, 2001 at 02:45:05, Dusan Dobes wrote: >On January 23, 2001 at 11:53:04, Robert Hyatt wrote: > >>On January 23, 2001 at 10:36:04, Uri Blass wrote: >> >>>On January 23, 2001 at 10:12:24, Robert Hyatt wrote: >>> >>>>On January 23, 2001 at 03:06:53, Dusan Dobes wrote: >>>> >>>>>On January 22, 2001 at 15:57:52, Dann Corbit wrote: >>>>> >>>>>>>I am interested if "your" engine could resist to play 46.Qe6+ in a blitz game. >>>>>>> >>>>>>>[D]6k1/2pq4/5Q2/p2PPP2/1p1P4/8/P6P/7K w >>>>>>> >>>>>>>Best wishes, >>>>>>>Steffen. >>>>>>> >>>>>>>[Event "ICS rated blitz match"] >>>>>> >>>>>>Phalanx has a time of it: >>>>>> >>>>>>[ white, 1 ] >>>>>>6k1/2pq4/5Q2/p2PPP2/1p1P4/8/P6P/7K w >>>>>> >>>>>>6k1/2pq4/5Q2/p2PPP2/1p1P4/8/P6P/7K w >>>>>> -> increment adds 1440 s to soft time limit >>>>>> -> soft time limit 3154.28 s >>>>>> -> hard time limit 41599.9 s >>>>>> 6 654 112 213089 Qf6-e6 Qd7-f7 Pf5-f6 Pa5-a4 Qe6-c8 Kg8-h7 >>>>>> Pe5-e6 Qf7xf6 Qc8xc7 Kh7-g6 >>>>>> 6 -> 0:01.60 305919 0 turns >>>>>> 7 654 583 1226489 Qf6-e6 ?? >>>>>> 7 352 742 1585651 Qf6-e6 Qd7xe6 Pf5xe6 Pa5-a4 Pd5-d6 Pc7xd6 >>>>>> Pe5xd6 Kg8-f8 Pd6-d7 Kf8-e7 Pd7-d8B Ke7xd8 >>>>>> Ph2-h4 Kd8-e7 >>>>>> 7 357 762 1623830 Qf6-g6 ! >>>>>> 7 525 919 1926808 Qf6-g6 Kg8-f8 Qg6-h6 Kf8-g8 Qh6-g5 Kg8-f8 >>>>>> Kh1-g1 Qd7xd5 >>>>>> 7 -> 0:10.26 2108193 1 turn >>>>>> 8 525 1161 2360687 Qf6-g6 Kg8-f8 Qg6-h6 Kf8-g8 Qh6-g5 >>>>>> 8 -> 0:23.94 4694659 0 turns >>>>>> 9 525 2935 5710337 Qf6-g6 Kg8-f8 Qg6-h6 Kf8-g8 Qh6-g5 >>>>>> 9 526 3916 7666834 Qf6-g5 Kg8-f8 Pd5-d6 Pc7xd6 >>>>>> 9 -> 1:10.73 13328419 1 turn >>>>>> 10 526 8747 16329005 Qf6-g5 Kg8-f8 Pd5-d6 Pc7xd6 Pe5-e6 Qd7-b7 >>>>>> Kh1-g1 Qb7-g7 Qg5xg7 Kf8xg7 Pe6-e7 Kg7-f7 >>>>>> 10 -> 3:58.68 42938008 0 turns >>>>>> 11 526 40425 71196909 Qf6-g5 Kg8-f8 Pd5-d6 Pc7xd6 Qg5-h6 Kf8-g8 >>>>>> Pe5-e6 Qd7-b7 Kh1-g1 Qb7-g7 >>>>>> 11 -> 17:32.80 184157117 0 turns >>>>> >>>>>So that's 7.62 seconds to switch to a different move. >>>>>Phalanx has special knowledge that helps it to see >>>>>those things quickly. It has huge extensions (like several >>>>>full plies) on piece exchanges leading to pawn-only endgames. >>>>> >>>>>Dusan >>>> >>>> >>>>I think you are overlooking the PV. IE it is _still_ showing a queen exchange, >>>>just not instantly. IE at depth=10. Trading on g7. Doesn't matter where >>>>you trade, white loses this if the b-pawn stays on the board. >>> >>>I think that programs have not time to extend enough in positions that are at >>>distance of 10 plies from the root. >>> >>>Practically it can see that Qe6 is not good at the root and I guess that in a >>>practical game it will also see 10 plies later the same problem. >>> >>>Uri >> >> >>Maybe not. IE it doesn't seem to understand that the b-pawn is going to >>win the game if queens come off. It doesn't matter whether you see it at >>the root or at the tips. You either have to search deep enough to see this >>problem, or realize that there are other moves that prevent it from happening, >>completely. > >You are right. It does not understand it in the static evaluation >function. I feel good detection of candidates is so difficult that >you have to mimic a search in the static eval anyway. I think using >extensions on right moves is better solution. what happens when you reach this position at the _end_ of your search? And you discover that you can either trade queens, or give up one pawn? If you trade queens, you lose no material but you are stuck with a forced loss you don't see. If you don't trade queens, your opponent picks up a pawn leaving you only 1 pawn ahead rather than 2. Which move do you choose then, since extensions won't help at a leaf position? The only reason I can't solve this with eval is that my eval has a passed pawn term that is dependent on remaining material, and the score gained by removing the queens (white passers become much stronger) more than offsets my "distant passed pawn candidate" evaluation score. If those white pawns aren't passed I can pick up this "win" (by removing the queens) instantly. But here taking the queen off boosts the score more than the opponent's b-pawn candidate. However, I believe that this kind of thing _must_ be caught by the eval. Because no amount of extensions will pick this up if it happens way out in the tree... > >> >>I'm going to file this one away as it is critical to get these right. I have >>the necessary code to detect this as won for black with no queens on, but I >>don't quite do it yet for safety reasons. But it has to be done eventually. >> >>IE if you back those white pawns up and block them with black pawns, Crafty >>will _instantly_ know to not trade queens as white, and will try to trade as >>black. But that isn't enough to offset the passed pawns white has, yet... >>yet being the operative word. :)
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.