Author: Christophe Theron
Date: 21:41:31 10/20/97
Go up one level in this thread
On October 20, 1997 at 13:58:37, Chris Whittington wrote: >Some programs mix and match between the two. They do swap-off >evaluations and terminate sometimes, or capture search sometimes, >depending on circumstances. Again the circumstances require knowledge >measurements that the fast programs can't/don't do. We started in "PV oscillation", and now we are in "quiescence search philosophy". Tiger uses a SEE to generate a list of all interesting (at first glance) captures, then really does the search to verify tactically. This saves a lot of time trying stupid capture moves. This pruning technique works well, and rarely does a mistake. But the SEE has to be quite complex. The reason to follow the interesting capture moves instead of trying to build a better SEE is that we have to evaluate the positional terms at the end of the exchanges (imagine a queen exchange for example). Maybe you could simply ignore the positional score if it seems to be well below alpha or well above beta... >Ed has reported his capture search is around 10-15% of total nodes. >CSTal's is also around this figure, maybe lower. It's not a constant value. In a very deep search it can fall below that even for a program that generates every capture. If the search is not deep, it can be more than 50% of the nodes. >I don't know the precise capture search rates for the fast programs, but >I think their rates are very much higher than these. Don't be so sure. That's one of the magics of alpha-beta :) - Christophe -
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.