Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: CT 15 and evaluation problem

Author: Tony Werten

Date: 11:15:46 11/05/03

Go up one level in this thread


On November 05, 2003 at 11:58:18, Christophe Theron wrote:

>On November 05, 2003 at 03:03:19, Tony Werten wrote:
>
>>On November 04, 2003 at 13:58:29, Christophe Theron wrote:
>>
>>>On November 04, 2003 at 02:43:34, Tony Werten wrote:
>>>
>>>>On November 03, 2003 at 13:22:42, Christophe Theron wrote:
>>>>
>>>>>On November 03, 2003 at 10:04:32, Will Singleton wrote:
>>>>>
>>>>>>On November 03, 2003 at 02:35:09, Gerd Isenberg wrote:
>>>>>>
>>>>>>>On November 03, 2003 at 02:33:02, Gerd Isenberg wrote:
>>>>>>>
>>>>>>>>On November 02, 2003 at 17:12:38, Will Singleton wrote:
>>>>>>>>
>>>>>>>>>On November 02, 2003 at 16:52:49, Gerd Isenberg wrote:
>>>>>>>>>
>>>>>>>>>><snip>
>>>>>>>>>>>In the same vein, the following position has always been a nightmare for Tiger:
>>>>>>>>>>>
>>>>>>>>>>>[D]8/1KP5/3q2k1/8/6p1/8/8/8 b - -
>>>>>>>>>>>
>>>>>>>>>>>This position comes from a real game between the Modular Game System Sargon 2.5
>>>>>>>>>>>and Mike III, played in September 1980 during the Personal Computer World Fair.
>>>>>>>>>>>Mike III continued the game with a long series of checks leading to a draw.
>>>>>>>>>>>
>>>>>>>>>>>Chess Tiger is not smarter than Mike III here. It is something that I had fixed
>>>>>>>>>>>in the 16 bits version, to the expense of some added complexity in the passed
>>>>>>>>>>>pawns evaluation code. I have not transfered this code to the 32 bits version
>>>>>>>>>>>because it was not general enough (add another black pawn and the code did not
>>>>>>>>>>>work).
>>>>>>>>>>>
>>>>>>>>>>>I am interested in results of other (amateur and commercial) programs.
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>    Christophe
>>>>>>>>>>
>>>>>>>>>>Hi Christophe,
>>>>>>>>>>
>>>>>>>>>>IsiChess on AMD XP2.6+ first shuffles around with Qb4+, Qxc7 and Qd7.
>>>>>>>>>>After 5 seconds at depth 13 Qxc7 came up. Mate in 14 resp. 12 after one minute
>>>>>>>>>>and 1:10. I guess a matter of won KPK eval.
>>>>>>>>>>
>>>>>>>>>>Cheers,
>>>>>>>>>>Gerd
>>>>>>>>>
>>>>>>>>>Sure, but I'm not sure CT was referring to finding the mate as the problem.  CT
>>>>>>>>>seems to have some problem understanding that KQK is better than KQKP, that's
>>>>>>>>>all I can figure.
>>>>>>>>>
>>>>>>>>>Will
>>>>>>>>
>>>>>>>you mean won KPK against KQPKP? I use interior node recognizers and assign
>>>>>>>shlightly more than queen advantage in won KPK. Additionaly there is a
>>>>>>>heuristic, that reduces score a bit (e.g. abs(delta material) / X) if a lot of
>>>>>>>checks occur with "no progress".
>>>>>>>                 ^^^^
>>>>>>>Gerd
>>>>>>
>>>>>>I do that too, but it's constrained to several consecutive checks that don't
>>>>>>reset the fifty-move counter.  Seems to help in some positions, but I'm not sure
>>>>>>how generally effective it is.
>>>>>>
>>>>>>Will
>>>>>
>>>>>
>>>>>
>>>>>That's exactly what I have tried, with different values for X and the number of
>>>>>consecutive checks needed to trigger it.
>>>>>
>>>>>It has always weakened my program. Not by much, but I expected a minor gain, not
>>>>>a minor loss!
>>>>
>>>>Make sure your trigger is counted from the back, not the front. ( I don't want
>>>>to make it too easy :)
>>>>
>>>>Tony
>>>
>>>
>>>
>>>I really don't understand what you mean...
>>
>>Whatever adjustments you do after no progress, it is only important if the last
>>fe 5 moves made no progress (or gave checks). If the first 5 moves did this and
>>then there is a no-check or capture you should do nothing.
>>
>>If you trigger from the front (ie root) you'll seriously weaken the engine
>>because sometimes it is nescessairy to shuffle some pieces around before doing
>>something, but you will be giving a penalty for that.
>>
>>Tony
>
>
>
>Oh yes. That's how I did it. When I reach a leaf I check if the last X plies
>where just checks without capture (X=10..12).
>
>Probably my definition of "no progress" was not restrictive enough. For example
>if there is a pawn push, it is not a case of "no progress" I guess.

I would call that a clear sign of progress :)

It's not going to make all horizon effects go away, but it does take care of a
few annoying ones. And it makes you look like a fool less often. I think I
decided to implement this after a draw with Gandalf at IPCC in a won position.

Tony

>
>
>
>    Christophe



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.