Author: Nagendra Singh Tomar
Date: 05:46:10 10/10/02
Go up one level in this thread
Thanx for the nice explanation.. I missed the point that move ordering is not perfect .. We can have a higher score move after a lower score move, though the lower score moves leads to promises, the higher score (> beta) move spoils all our hopes .. regds tomar On October 10, 2002 at 08:17:10, Carmelo Calzerano wrote: >On October 10, 2002 at 07:16:27, Nagendra Singh Tomar wrote: > >>In ur example; you should note that when we are searching for the opponent's >>best reply for d4 we always have alpha and beta (calculated till the first >>subtree e5 is fully evaluated)in hand. If we find any of the opponent's reply to >>be better than e5 (and asusming that e4 was THE best move), the score of the >>reply WILL be more than the beta set by e4 hence it will fail-high and we will >>not update PV. IF at any point in time we get a move that lies between alpha and >>beta we know that the move lies on the PV based on the tree calculation till >>that time. >>IOW if at any point in ply 'x+1' we update the PV, this updation WILL percolate >>all the way upto the top, updating the PV moves of all the plies <= x. > > >That's a wrong assumption. Think a minute about it: even if, at ply N, >your current search score lies between alpha and beta, you cannot possibly >assume that at ply N-1 a move not yet analyzed will cause a fail high, >rejecting the move at ply N-2. For instance, just add one ply at your >own example and it won't work anymore: > >The current PV is 1. e4 e5 2. Nf3, and we are now analyzing 1. d4 d5. >We try 2. c4 at first, and we find a score lying in the alpha-beta window. >Thus, our PV become [1. d4] d5 2. c4 and we lost all the information >about the move 2. Nf3. Note that the analysis of 1. d4 is not finished >yet, so in our PV array we actually have [e4][d5][c4]. > >Now we analyze 1. d4 Nf6, and let's suppose Nf6 fails high: the 1. d4 >move is rejected. The correct PV should be still 1. e4 e5 2. Nf3, but >you actually lost it: your PV array is in an unconsistent state! > >That's why you CANNOT assume that a score lying between alpha and beta >will propagate up to the tree, because at the tips you have no idea >how many moves are left that could possibly reject the whole variation. > > > >>beta is there to say that the opponent has got means to stop u from playing a >>move, so score>beta cannot lie on PV. For a move for which alpha < score < beta, >>the opponent has got no means to stop us from playing this move. > >Unfortunately that's not true... ;-) > >Regards, >Carmelo
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.