Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Book Learning

Author: Thom Perry

Date: 05:00:37 12/31/98

Go up one level in this thread


On December 31, 1998 at 02:13:55, greg moller wrote:

>On December 30, 1998 at 09:13:52, Dan Kiski wrote:
>
>>On December 30, 1998 at 07:25:29, Thom Perry wrote:
>>
>>>On December 30, 1998 at 05:38:41, greg moller wrote:
>>>
>>>>On December 30, 1998 at 03:39:09, Komputer Korner wrote:
>>>>
>>>>>On December 29, 1998 at 16:20:06, Dan Kiski wrote:
>>>>>
>>>>>>The following game is from Fritz 5.16 v MCP8
>>>>>>
>>>>>>Fritz 5.16 white
>>>>>>MCP8       black
>>>>>>
>>>>>>1.  NF3......NF6
>>>>>>2.  D4.......E6
>>>>>>3.  C4.......B6
>>>>>>4.  NC3......BB7
>>>>>>5.  A3.......D5
>>>>>>6.  CD.......ED
>>>>>>7.  QA4 +....C6
>>>>>>8.  G3.......NBD7
>>>>>>9.  BG2......BD6
>>>>>>10. 0-0......0-0
>>>>>>11. BF4......QE7
>>>>>>12. BXD6.....QXD6
>>>>>>13. RFD1.....RFD8
>>>>>>14. E3.......A5
>>>>>>15. RAE1.....BA6
>>>>>>
>>>>>>Fritz is now out of book but chooses NH4, MCP8 is still in book expecting QB3.
>>>>>>
>>>>>>Now the postion after 16. NH4 i have played through countless times at varying
>>>>>>time controls yet MCP8 seems to always choose 16.......NG4.
>>>>>>
>>>>>>It has lost from this position over and over but always chooses 16.....NG4 and
>>>>>>after Fritz's 17. E4 replys with 17......QF6 it might from this point try to
>>>>>>change the odd move yet according to Fritz it was already way in front, can
>>>>>>someone explain how this book learning works and how many losses will it take
>>>>>>before the computer in this case MCP8 varies it's responce, in this case most
>>>>>>likely remove 16........NG4 since this seems to be the start of its problems.
>>>>>>At one point it added 16......NG4 to book after 6 straight losses using it.
>>>>>>
>>>>>>I wait until it is 7 points behind claim the win and then scroll back. Play 16.
>>>>>>NH4 then carry on again. Am I doing something wrong with this method.
>>>>>
>>>>>Maybe MCP8 doesn't learn if you claim a win.
>>>>>--
>>>>>Komputer korner
>>>>
>>>>Well, it would be interesting to test this hypothesis, since somehow i always
>>>>had the impression that the older MCP7 was adding book moves in the lines it was
>>>>LOSING with (!) At this point in time I'm suspicious about this whole learning
>>>>algorithm. How exactly does it work, and can it be verified ?
>>>>
>>>>Regards, greg
>>>
>>>I too have been wondering about the learning feature in MCP8.  I remember back
>>>in MCP4/5 that once you analyzed a position/game, if you wanted to repeat the
>>>original results, it was necessary to turn off the learning features.
>>>Otherwise, the second time around, the evaluation score(s) and response time(s)
>>>(time it takes for the program to make a move) were usually slightly different.
>>>I attributed this to the learning features.  However, with MCP8, I have played
>>>several positions where the evaluation score changes drastically (downward)
>>>after MCP8 has selected a move, yet MCP8 will continue to make the same move,
>>>over and over.  This doesn't look like learning to me.
>>
>>I have no doubt that I do have it set correctly I have been through the manual
>>countless times, my only doubt was when I claim the win when 7 pts in front then
>>scroll back if it accepts that and learns from it, from what I have seen is it
>>accepts it since as stated at one point it added 16. NH4.....NG4 to book, yet
>>don't learn from it. Maybe 16.....NG4 is ok maybe the problem is with
>>17........QF6 in reply to 17. E5.
>>
>>Yet as of yet I haven't been able to get it to make another move at time
>>controls of up to 3 min\move it just keeps walking into losses even with Fritz
>>5.16 set at 10 secs\move from this position.
>
>  I hope that Marty Hirsch, the programmer, gets wind of this problem somehow.
>What I find odd is that I read in KK's(I think) review of MCP8 that book
>learning had supposedly been improved (!) , this according to Marty, and now it
>boasted the best such learning feature. Then there was a report (in CCC?-not
>sure) by someone who tested Nimzo98 vs MCP8 and noticed that Nimzo's learning
>feature was superior. This cast further doubt on the MCP8 claims.
>
> Unfortunately, until there's more rigorous testing done, we won't be able to
>take these claims for granted. IMO if this is a bug it should be fixed ASAP,
>since Mchess is a really fine program.
>
>
>regards, greg

I copied this thread and sent it to Marty this morning.  One of the reasons I
have been buying from him since MCP4 is because he does provide good support for
his program.  He puts his e-mail address in the owner's manual, and he has
always (with only one exception!) promptly responded to my questions about MCP.





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.