Author: Kristo Miettinen
Date: 15:23:57 05/06/99
Go up one level in this thread
Hi Jose! A different idea that I tried once was to vary my evaluation function parameters within each iterative-deepening search to try to impose consistency from iteration to iteration. In other words, not training on grandmaster examples but rather striving to "see" at depth N-1 what you found at depth N, for the specific position at hand. The pseudocode goes something like this: Begin with initial evaluation function from previous search or from a database. Set search depth N = 2. Label A: Conduct depth N search. Remember results of depth N-1 search with same parameters (no computation). Vary an evaluation parameter (I used an exhaustive sequence of parameters and fixed small adjustments). Conduct a depth N-1 search with the modified evaluation function. If the modified evaluation agreed more closely with the depth N search (meaning chose the same move and gave a closer value estimate) than the depth N-1 search with the unmodified evaluation function, then accept the modification, check time consumption, and if enough time remains return to Label A above (otherwise exit and move). Make the opposite variation to the evaluation parameter (if incrementing a parameter made things worse, then maybe decrementing it will make things better). Conduct another depth N-1 search. If the modified evaluation agreed more closely with the depth N search then accept the modification, otherwise return to the original evaluation function. If the original evaluation function has been retained after both attempted modifications, increment N. Note that changing the evaluation function because of the results of a depth N-1 search causes a new search at depth N, not at N+1. Check time consumption, and if enough time remains return to Label A above, otherwise exit and move. I gave up on this approach when I gave up on iterative deepening and went instead to an architecture with a connected graph of the entire search stored all in memory, with the graph grown by selectively expanding one node at a time. Sine cera, -Kristo.
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.