Author: Tord Romstad
Date: 04:05:23 08/03/98
Go up one level in this thread
On August 02, 1998 at 07:50:08, Robert Hyatt wrote: >On August 01, 1998 at 14:15:33, Robert Pawlak wrote: > >>I was wondering whether Rebel 9's time usage is realistic when operated in the >>ELO handicapping mode. >> >>I ask this because Fritz (my mainstay) does not have realistic time usage when >>in the ELO modes. It seems to move instantaneously when set at 1800-1900. >> >>Both CM5500 and CStal seem to have good/realistic time usage when playing at the >>weaker levels. Can rebel do the same? >> >>Thanks, >> >>bob P. > >There are at least three well-known ways to "dumb down" a program: > >1. reduce the search time (a time-handicap mode) to give the opponent >more time. This has to be done in conjunction with disabling "think on >opponent's time" or it won't work well. > >2. reduce the positional scoring terms so that the program loses the >significance of positional considerations. IE passed pawns become less >valuable, open files become less important, pawn structure becomes less >important, ditto for king safety, etc. > >3. factor in some sort of random number so that positions that are bad >will randomly look less bad or actually look good. Or, if the random >addition is large enough, but infrequent enough, the program might begin >to make tactical blunders (ICC has some *bach programs that do (or did)) >this, and they would make an occasional blunder like a weaker human). I have considered the following idea: Make the move generator "overlook" a few moves. Keep a counter of the total number of moves the move generator has generated during the search. When this counter reaches a multiple of some fixed number n, you simply avoid searching the next move the move generator returns. This would make the program overlook tactics at random places in the search. The number of tactical oversights can be controled by varying the number n. This idea could be refined in several ways. For instance, you could try to make the oversights more "realistic" by increasing the probability of oversights deep in the search, or by removing only moves which seem to lose material (if a swap-off function is used to order the moves, this should be trivial to implement). I haven't yet implemented this feature in my program, however. Has anybody else experimented with similar ideas? Tord
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.