Author: scott farrell
Date: 13:30:16 02/08/03
Go up one level in this thread
On February 08, 2003 at 10:34:07, Robert Hyatt wrote: >On February 08, 2003 at 05:13:11, scott farrell wrote: > >>After all the ideas I have 'borrowed' from crafty, I would only hope that one >>day I could be responsible for a line or 2 when crafty plays the next CCT. >> >>SCott > > >I'm pretty sure this is going to produce worse results in Crafty. A couple of >things: > >1. You are using a mixed-depth search, D+1 for the first root move, D for >the rest. I know of no way to _compare_ those results. IE if a D+1 fails >high, and I don't have time to re-search it, what do I do? I can't trust a >D ply fail high over a D+1 true score, because the D ply search is shallower >and could be overlooking something critical. Robert, firstly thanx for looking this over. I am using ply and ply-1, which is essentially the same as ply+1 and ply, so I'll try not to confuse the conversation too much. I dont compare D+1 ply results with D ply results. During the PVS search of the remaining moves using ply-1, if they fail low during PVS obviously they are discarded, if they fail high during PVS, I research (as normal with PVS) but with ply (instead of ply-1), and only then do I compare the results. I never use the move from the search at ply-1. The only risk of oversight is there is a better move in the ply-1 search if I searched deeper to ply. In your comment above "IE if a D+1 fails high, and I don't have time to re-search it, what do I do?" Given ply+1 was the PV move, more often than not the first move after sorting at root, if it fails high yipee, it means it is more the previous iter, I'm pretty sure its atleast safe if not a winner - if you dont research becuase of time on a fail high there is no risk just take it. > >2. You have to watch the hash table. Else you are going to store an entry with >the wrong depth at the root. That will certainly cause a problem on the next >iteration. Yes, been there before, I think I have the implementation correct in this regard. thanx again Scott
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.