Author: J. Wesley Cleveland
Date: 10:15:31 08/06/02
Go up one level in this thread
On August 06, 2002 at 06:12:46, Sune Fischer wrote: >On August 06, 2002 at 05:32:11, Gian-Carlo Pascutto wrote: > >>On August 05, 2002 at 19:45:57, James Swafford wrote: >> >>>On August 05, 2002 at 18:01:57, Vincent Diepeveen wrote: >>> >>>>On August 05, 2002 at 17:58:16, James Swafford wrote: >>>> >>>>this is in my alltime CCC posting. >>>> >>>>If a value from +3 goes to -3 then that's the most absurd thing >>>>which can happen. >>> >>>Of course, if you're implying -3 means "it's good to get rid of >>>the bishop" or some other similar nonsense. But the point is >>>that moves are more important than eval term weights. >>> >>>Surely you agree... (?) >> >>You're presenting it as if they are two independent things, which is >>surely not true. > >The idea is to give the computer free hands to optimize in whatever way results >in the most profitable play. > >Vincent gave an example where it was too agressive giving a piece for 2 pawns >and a check. My guess is, that this over-agressive (patzer) play does result in >better overall score for the program. Yes it will be destroyed from time to >time, but if there is still a net gain, then that is what the algorithm will aim >for. > >IMO it is possible to correct this by simply giving it the right opponents while >it trains, it must be punished for being over-agressive. > I'm having difficulty with this idea of "right opponents." If the opponent is considerably stronger, then over-aggressive play is probably correct, as sometimes the sacrifice will be correct and it will win, while it will always lose without it. If the opponent is equal or weaker, then often bad moves will be rewarded as the opponent will make even worse moves. An idea that just occurred to me is repeatedly self-playing against a version with one less evaluation term. When the weights for that term settled, make that the base version and add a new term to the test version.
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.