Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Fine tuning the engine's strength

Author: KarinsDad

Date: 07:52:40 07/25/00

Go up one level in this thread


On July 25, 2000 at 05:26:58, blass uri wrote:

>On July 25, 2000 at 01:26:15, KarinsDad wrote:
>
>>On July 24, 2000 at 20:26:27, John Coffey wrote:
>>
>>>On July 24, 2000 at 14:45:01, KarinsDad wrote:
>>>
>>>>On July 24, 2000 at 14:23:19, KarinsDad wrote:
>>>>
>>>>>On July 24, 2000 at 13:30:06, Jari Huikari wrote:
>>>>>
>>>>>>On July 24, 2000 at 13:01:36, John Coffey wrote:
>>>>>>
>>>>>>>Only slightly related to the GUI is having a range of abilities from beginner
>>>>>>>up to the top level that can be fine tuned.
>>>>>>
>>>>>>>I tried it on Chessmaster 6000, all the levels 1600 and below were dropping
>>>>>>>pieces, and the next level up was smashing me at speed chess (my quick rating
>>>>>>>is 1978.)
>>>>>>
>>>>>>I have thought about how this could be done. One idea that came into my
>>>>>>mind was simply to put some delay routine into search to make it slower
>>>>>>and thus playing weaker.
>>>>>>
>>>>>>					Jari
>>>>>
>>>>>
>>>>>I do not think those types of solutions work, i.e. less time, fewer nodes, lower
>>>>>depth, etc. The program will still play relatively strong until some other
>>>>>algorithm takes over (i.e. the below 1600 drop piece problem that John noted).
>>>>>
>>>>>What you need is a chess engine that generates multiple ply 1 PVs. Then, it
>>>>>could randomly pick a different PV each move.
>>>>>
>>>>>So, for example, if it had 5 PVs that it could choose from, at 2600 setting it
>>>>>would always pick PV 1 each time. At 2400 setting, it would occasionally pick
>>>>>the PV 2 move. At 2200, it would pick PV 1 45%, PV 2 45%, PV 3 10%. At 1600, it
>>>>>might pick PV 1 20%, PV 2 20%, PV 3 20%, PV 4 20%, PV 5 20%.
>>>>>
>>>>>Then, the computer would not be dropping pieces, even at a 1000 setting (even
>>>>>though 1000 players often do drop a piece). But, it would rarely be playing the
>>>>>best move in those positions at the lower settings.
>>>>>
>>>>>Of course, you would have to add in some logic that the scores of the PVs could
>>>>>not be that drastically different. For example, NxB would normally result in PxN
>>>>>as PV 1. If PV 2 did not have a similar PV score to PV 1 (i.e. there were no
>>>>>waiting moves that do not lose the bishop), then the program would still make
>>>>>the PV 1 move, regardless of setting.
>>>>>
>>>>>KarinsDad :)
>>>>
>>>>I forgot to mention that lowering the depth in conjunction with this type of
>>>>solution would be optimal. It doesn't make sense to pick a PV 5 move that avoids
>>>>a capture 14 ply down that is also avoided by PV 1 through 4. If the setting is
>>>>1200 rating, then the program should not generally be seeing more than 4 to 6
>>>>ply down before deciding on it's PVs.
>>>>
>>>>KarinsDad :)
>>>
>>>
>>>The idea has occured to me before.  Write a program that would decide randomly
>>>between between the best N # of moves where N is based upon the level of
>>>difficulty.  I saw the same problem that you saw which was that sometimes
>>>the best move is forced.  So then you have to decide how much of an error you
>>>will allow.  If you will allow a 3 point error at 1300 but only a 2 point
>>>error at 1400 then you have the problem that 1300 will drop knights and 1400
>>>won't.  Maybe that isn't a problem.  But maybe there would be too much
>>>difference between 1300 and 1400.
>>>
>>>John Coffey
>>
>>I think the issue is one of having a program attempt to decide the average
>>"power" of a move based on the setting desired. Once that is decided (does that
>>mean only searching 8 ply, does that mean randomly picking between the perceived
>>best three moves, etc.), the program can still play reasonably well, but it can
>>also blunder if a piece hangs in 6 ply, etc.
>>
>>The problem comes in when the program makes a move that is SO stupid that not
>>even a low rated player would make that move (in standard time controls, in
>>blitz, almost everyone but extremely good players blunder). When this happens,
>>the program is considered a dumb machine.
>
>Even high rated players do sometimes blunders of not seeing one move.
>It does not happen often but it happens.
>
>I remember that it happened even to one of the world champions.
>
>Uri

Yes, Fischer missed mate in one.

KarinsDad :)



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.