Author: fca
Date: 16:32:32 04/13/98
TRY TO MAKE LIFE HARD FOR THE OPPONENT! Where several (2 or 3) moves keep shuffling about as being best-so-far in various successive plies, they probably are almost indistinguishable.... So (especially if the enemy is DNA-based) make the final selection between the two *not* with the ply-dice (i.e. "at this depth does this move displace the other one") but as follows: 1. Use a tree-audit perhaps on a sample basis. 2. Choose between the candidates *not* on the basis of their backed-up evals but instead on how *few* lines were reasonably successful against each... So, if we are white, and + is good for us, and the candidate moves (swapped around a few times) are say w1 and w2; if evals (obtained by excursion) are: w1 b11 -- +1.0 w1 b12 -- +1.1 w1 b13 -- +1.4 w1 b14 -- +1.2 and w2 b21 -- +1.4 w2 b22 -- +0.98 w2 b23 -- +1.8 then I would choose w2 even though it minimaxes to 0.98 < 1, as the bonus if black misses b22 is *big* ( i.e. 0.4 >> 0.02) (0.4 = b21 - b11; 0.02 = w1 - w2). If getting backed-up evals is too slow, sample the Black responses including previous killer candidates etc..... Yes I know 0.98 may not be an available number - hence I suggest the trigger mechanism for us to go into this process is the swapping around of candidates as search deepens. As comps get it wrong too, so this should not only be an anti-human style, though kick-in values could be set according to oppo type: if the perm brain has correctly guessed a high proportion of oppo moves, it is probably a comp) Chinook did it yonks back when in its endgame databases and having to distinguish between non-winning lines.... I propose extending it to incomplete information datafields i.e. all of chess. ICCAJ article (c10yrs back) " Pathology in Game Trees" is also interesting.... Ideas? Comments? The process is time-cheap, and at worst it has done nothing useful. Shouldn;t have screwed up the hash table either.... Kind regards fca PS: And I can't find a trace of this in crafty - I looked :-)
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.