Author: Uri Blass
Date: 14:29:17 04/01/03
Go up one level in this thread
On April 01, 2003 at 16:22:33, George Sobala wrote: >On April 01, 2003 at 14:07:45, Robert Hyatt wrote: > >>For a computer to almost break even at 3 0 is _terrible_. > >Robert, I really think that you continue to miss the point of the redshift >settings for Deep Sjeng. Of COURSE they cripple the engine's chess playing >skills - but not in the usual ways of reducing the search depth or making it >deliberately blunder once in a while. Instead, the redshift personality has a >WILDLY unbalanced view of the relative values of pieces on the one hand, versus >mobility and king attack on the other (so that e.g. it may have an eval of +5 to >+10 just for an attack). And I mean wildly - even Gambit Tiger on "suicidal" >setting is a mere pussycat compared to redshift's crass stupidity in attack. > >The interesting consequence, is that whilst other comps eat it for lunch (e.g. >even mscp on ICC rated around 1700 and with a search depth of 5 will regularly >beat it), humans up to (weak) IM level struggle to do so. (I grant you that I >merely refer to performance at ICC-type time settings.) I find it interesting >that its loss of playing skill against humans is disproportionately small >compared to its huge loss of playing skill against comps. > >Anyway, the net result is an engine which will give a weak human player a fun >game, where Bxh6 or Nxf2-type sacs by the comp do not automatically mean "Oh >dear, I may as well resign". > >And the hypothesis is that perhaps the strongest chess engines may challenge >humans better in the future by being tuned to the weaknesses in human play, by >being more adventurous and unsound, even whilst compromising their performance >against other comps. I think that it is a bad idea to play unsound sacrifices if the target is increasing the playing strength against humans. Computers already beat GMs without it and if the target is to win against the best humans then unsound sacrifices are a bad idea because the best humans have good chances to defend correctly and win the game. I understand that your claim is that the difference between it's rating against humans and it's rating against computers is more than 400 elo for it's rating against humans. My opinion is that it suggests that humans simply did not try to learn the right choice of opening to play against it. It may be interesting what happens if humans play to win and not for fun. I suspect that the difference will be smaller because humans may go for positions when it is harder to find sacrifices so Sjeng will have to try something more crazy for sacrifices and it will do it and with bigger advantage it is going to be easier for humans to win. Uri
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.