Author: José de Jesús García Ruvalcaba
Date: 10:01:41 07/19/99
Go up one level in this thread
On July 18, 1999 at 19:31:23, Francis Monkman wrote: > >On July 18, 1999 at 16:05:00, Robert Hyatt wrote: > > >>Here's why it has a problem. Take a typical middlegame with roughly 35 moves. >>If you have 1000 computers, you can go two half-moves into the future and search >>those 1000 (35*35 roughly) positions. However, this is not very efficient as >>alpha/beta needs the score for the 'best' (first) move before searching any of >>the other moves... >> >>it is very difficult... > >I guess we could be talking more than that. Some guys have been trying to work >out from the voting % the minimum number participating, but I've heard said >250,000 -- no idea if that means anything. Anyway, if they're voting, they're >online someway. > >Couldn't we figure something out to do full-width for the first few ply or >something? That would already be quite something, especially as the alpha-beta >algorithm seems to have some problem here (ie good opponent moves not searched >-- this topic's been discussed)? > >In fact, this game's showing up some interesting bugs in Fritz 5.32's >calculations. There's a particular type of position in this game (in some >lines), where Fritz wants ...Kc8, searching for minutes -- but the moment it's >played, the eval goes bad by around 1 pawn! Fritz is a root processor, the behavoir you see (big drops in the eval after a move is made) is normal in root processors. That means that the evaluation of a leaf node depends also on the root position. When a move is made, the root position changes and with it the evaluations of all its sons in the search tree. This is the drawback of root processing, the advantage is that it can be much faster. Alpha-beta will never cause those problems by itself (unless there is a bug, of course). Alpha-beta is a "safe" pruning, it produces the same best move as searching all the posible moves to the same depth (but it is faster, I think it looks at approximately the square root of the number of posible positions). José.
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.