Author: Uri Blass
Date: 01:54:16 07/17/02
Go up one level in this thread
On July 16, 2002 at 23:30:40, martin fierz wrote: >On July 16, 2002 at 20:31:49, Robert Hyatt wrote: > >>On July 16, 2002 at 18:39:53, Gerd Isenberg wrote: >> >>>On July 16, 2002 at 18:25:08, Robert Hyatt wrote: >>> >>>>On July 16, 2002 at 18:10:59, Uri Blass wrote: >>>> >>>>>On July 16, 2002 at 17:44:36, Robert Hyatt wrote: >>>>> >>>>>>On July 16, 2002 at 17:38:02, Uri Blass wrote: >>>>>> >>>>>>>On July 16, 2002 at 16:52:33, Robert Hyatt wrote: >>>>>>> >>>>>>>>On July 16, 2002 at 14:32:57, Russell Reagan wrote: >>>>>>>> >>>>>>>>>On July 16, 2002 at 11:07:10, Robert Hyatt wrote: >>>>>>>>> >>>>>>>>>>However, with _no_ book you leave yourself open for preparation of traps. I >>>>>>>>>>saw Ken Thompson do this to NuChess years ago at an ACM event. You don't want >>>>>>>>>>to leave that kind of "hole" for a major event... >>>>>>>>> >>>>>>>>>Was it a trap set specifically for that engine? Or was it just a general trap >>>>>>>>>that many engines fall for when left on their own? >>>>>>>>> >>>>>>>>>Russell >>>>>>>> >>>>>>>> >>>>>>>>Sort of both. >>>>>>>> >>>>>>>>1. Ken knew which opening Nuchess would play, as ken was white and they >>>>>>>>had a pretty narrow book. He simply added a line that made them go out of >>>>>>>>book pretty early, with a classic trap where they gave up a piece to win >>>>>>>>the rook at a1, and thought they were winning an exchange and losing a pawn. >>>>>>>>In reality, they lost both pieces plus the pawn, and the game. >>>>>>> >>>>>>>It is not going to work against Movei even in bullet. >>>>>>> >>>>>>>Movei evaluates black knight at a1 as clearly less than a knight in normal >>>>>>>squares and the difference in evaluation is slightly more than a pawn. >>>>>> >>>>>>Look out on ICC. Would you rather lose a pawn or be forced to place your >>>>>>knight on A1/H1 for a while? I would prefer to stick the knight on the >>>>>>corner, because that can be corrected later. Losing the pawn is losing the >>>>>>pawn, period. You can't make it come back later. >>>>> >>>>>There are often cases when the black knight cannot come out of a1 later. >>>>> >>>>>I prefer to lose a pawn for a probability of 40% to win the knight and I suspect >>>>>that it is more than 40%. >>>>> >>>>>If the knight is not trapped there is a good chance that the search can get it >>>>>out of the corner and it is easier relative to the case that the knight is >>>>>trapped because if it is trapped you may need many moves to capture it and in >>>>>most of the cases when it is not trapped you can get it out of the corner in one >>>>>move. >>>>> >>>>>I also saw games when Movei beated other programs because the knight of the >>>>>opponent was trapped in the corner. >>>>> >>>>>I think that these cases are more common then the cases when Movei is losing >>>>>games because it evaluates the corner too much. >>>>> >>>>>Uri >>>> >>>> >>>>I used to have a much larger corner penalty. And found on a few occasions >>>>that playing Nh1 to defend the pawn on f2 was the right move. But the program >>>>refused to do so because it thought that having the knight on H1 was worse than >>>>losing the pawn on f2. It was unfortunately mistaken. :) >>>> >>>>Big positional scores are perfectly OK. But you had better be sure that they >>>>don't have outlandish side-effects. This one sounds like it will... at a time >>>>when you least expect it. >>> >>>Did you mean a white knight on H1? >>>May be the corner penalty makes more sence with white pawns on A8 or H8 only, >>>where they are likely to get trapped. >> >>I thought Uri said "corner squares". Which is more dangerous than "corner >>squares on the enemy side of the board". But even then there are probably ways >>to exploit a program that thinks a knight on a8 is really bad... > >like when it's good to go and grab the exchange the program won't do it... >sounds to me like not evaluating it and evaluating a knight in the corner as -1 >pawn are both quite wrong - a bit more knowledge should be able to do much >better than that! > >aloha > martin Evaluating knight at the corner as more than -1 is knowledge. We do not know the best evaluation and it is an open question but saying that both things are wrong is not productive. I do not like criticizing everything and saying that more knowledge is needed. I believe that there are simple changes that may make the program better and it is better to think about simple changes before adding complicated things. Adding knowledge to the evaluation is also relatively less simple in the case of movei because I do not use only the leaf position for my evaluation. When I look in my code I see that I evaluate also knight at the corner that is not the opponent corner as more than -1 (2.20 for white knight at a1 and 2.00 for white knight at a8 when the normal value of knight is usually close to 3.50 pawns). I do not remember cases when movei lost because it sacrificed a pawn for knight in the corner so I do not see it as a big practical problem. White knight at a1 is not going to be trapped often but the passivity of the knight can be worth more than a pawn. If the knight is not passive the search is going to get it often out of the corner. Uri
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.