Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: morph(C) on ICC

Author: Robert Hyatt

Date: 15:12:57 10/12/99

Go up one level in this thread


On October 12, 1999 at 10:02:54, Bas Hamstra wrote:

>Not that it's form of learning comes close to it, but KnightCap learned itsself
>from 1600 to 2200 in a couple of months, according to the author. It adjusts
>various eval paramaters based on the result of the game. It thinks it's pretty
>interesting.
>
>


yes, but that was _totally_ different.  Knightcap had an eval, it just
learned by adjusting the various weights.  Morph has _no_ eval of any kind
to adjust.  It build _everything_ as it goes.  All it knows is the actual
rules of the game, including how the pieces move and all about checkmates
and draws.  Everything else it has to learn on its own, including piece
values...

very different.  very difficult.  very ambitious.




>Regards,
>Bas Hamstra.
>
>
>On October 12, 1999 at 09:24:53, Robert Hyatt wrote:
>
>>On October 11, 1999 at 20:31:47, Manuel Rodriguez Blanco wrote:
>>
>>>On October 11, 1999 at 18:07:12, Robert Hyatt wrote:
>>>
>>>>On October 11, 1999 at 16:12:27, Jeff Anderson wrote:
>>>>
>>>>>On ICC there is a chess program called morph(C).  Does anyone know anything
>>>>>about it?  Here are it's notes and stats.  It has amzingly low ratings.
>>>>>
>>>>>Information about morph(C) (Last disconnected Fri Oct 08 1999 06:26):
>>>>>
>>>>>          rating [need] win  loss  draw total   best
>>>>>Bullet       467         19   398     7   424    564 (13-Sep-1999)
>>>>>Blitz        557        103  1269    22  1394    978 (11-Sep-1999)
>>>>>Standard     754         55   553     9   617    935 (06-Feb-1999)
>>>>>
>>>>> 1: Morph IV: A Learning Program
>>>>> 2: Morph is a project at UC Santa Cruz
>>>>> 3: Operated by 'weber' (UCSC Morph Team)
>>>>> 4: Currently Running on a Sparc Ultra-10 and Solaris 2.7
>>>>> 5: Blitz and rating < 1200 unless i am learning ( mood = 1 )
>>>>
>>>>
>>>>this is a program that only knows the legal moves of chess, and nothing else.
>>>>It has learned how much the pieces are worth, and everything else, by rules
>>>>that are generalized internally.
>>>>
>>>>Whether this approach can work or not is anybody's guess...
>>>
>>>Hi, i find this extremely interesting, do you know something more about this
>>>programs or others like this?, can you put an example of this rules that are
>>>generalized internally, what are your personal opinion about this type of
>>>programs?  do you think that type of program that learn, will be more common in
>>>the future?
>>>
>>>thanks in advanced to Hyatt or other that can answer some of this question or
>>>can give more information about this type of programming and this type of
>>>program
>>>
>>>Manuel
>>
>>
>>All I can tell you is from personal experience.  At one of the ACM events,
>>they had this thing running on a Sun in the back of the room.  I played it
>>several very fast games (human, not using Cray Blitz).  In the first game,
>>the game ended in a sort of 2 knights, with my bishop on c4, knight on
>>g5, and it didn't understand so it let me play Nxf7 forking Q and R.  When
>>we alternated colors, it played Nf3 Ng5 Nxf7 forking my queen and rook.
>>Of course, I played Kxf7 as the bishop wasn't supporting the knight.  But
>>it 'learned' that forking two pieces was good.  And it eventually figured
>>out that the knight has to be supported by a B, or else Bxf7 is no good.
>>
>>I didn't see any great promise, but it definitely had a form of generalizing
>>learning...  that is all that I know, beside the fact that it did not have any
>>sort of alpha/beta search component...



This page took 0.02 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.