Author: Walter Faxon
Date: 23:54:23 04/07/05
Go up one level in this thread
On April 07, 2005 at 10:38:41, Charles Roberson wrote: > > You misunderstand. Modern programs are not truely "full width" searchers. > They prune 90% to 95% of the positions. Where the branchfactor is near 40. > The effective branch factor is near 3. The search is a directed search > thus it only evaluates what it needs to. > > Once the PV is established, move ordering is used to create fast cutoffs. > Thus, modern programs only search what they need to (within their > understanding) and evaluate only what they search. Now, there is Lazy > Evaluation which prunes some of the evaluation effort in very unbalanced > positions. Thus, current practice is not to do full evaluation on all > positions evaluated. > > The way your statement goes (if I understand it), you assume a full width > search is done. That is not truely the case. > > Today, the phrase "full width" means that it is provable that this search > produces the same results as a full width search without being a true full > width search. That is where the statements of "safe pruning" and "unsafe > prunning" come in. Sorry, I meant to say "brute-force" search. Even with the best forward-pruning methods, which find tactics to prune out blunders and extreme time-wasting, most frontier positions are unlikely, even ridiculous. But most of these positions are adequately dealt with by static evaluation. I guess I'm assuming (dangerous!) that we can someday develop a test for determining how "interesting" -- if not "hard" or "critical", my prior terms -- a position is. Presumably we will want to put a greater effort into investigating these positions. My question is how to reduce the total search effort by delaying such investigation, since other, easier to evaluate, parts of the tree might obviate it. Here's an exercise to see if your forward-pruning heuristics are hurting your search: From a sample of games between strong opponents (human or computer), have your program analyze each position. You want the actual moves played to appear in your tree; the deeper the better. If not, what is removing them? Either way, good luck with NoonianChess! -- Walter P.S. I used to work with a fellow named Joe Noonan... :)
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.