Author: Michael Yee
Date: 10:15:36 12/14/05
Go up one level in this thread
On December 14, 2005 at 09:39:05, Charles Roberson wrote: > > A post yesterday put me on a path. In simulated annealing > one allows the system to backup a bit. This has proven to be > better than a purely deterministic approach. So, why not the > same thing in a chess search? > > At each comparison point, allow the search to take on a > new best move even if the score is lower. Here one would allow > a small range (obviously much less than a full pawn). Maybe, it > would have a similar effect on the search as it does in simulated > annealing? > > Thoughts: > a)Debugging would become more of a pain. > b)It would add a level of nondeterminism which could be nice. > c)I know in my case (maybe others believe the same), that the > position evaluator is not completely accurate in the range > 1 centipawn maybe even 30 centipawns. This thought adds to > the interest of adding a simulated annealing type feature. > > What say those of you out there? > > Charles In computer go, there are some attempts at using a simulated annealing type approach: Brugmann (Bruegmann?), "Monte carlo go" Bouzy, "Monte carlo go developments" However, these are more in the expected-outcome realm than the minimax realm, (see a reference from Abramson on EO). Sorry I don't have more precise references handy. But google/google scholar/citeseer should help out. :) Michael
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.