Author: Vincent Diepeveen
Date: 21:37:14 05/05/01
Go up one level in this thread
On May 05, 2001 at 08:25:56, Uri Blass wrote: >There is a possibility to set selectivity of Deep Fritz to 0 and it means that >Fritz avoid null move pruning. Except for some tactical shots it shouldn't matter much for a program like fritz which searches in a very sound way (no dubious pruning, because if you combine that with hashtable and nullmove then all hell breaks loose). At small depths however a small tactical gain might matter, so the fullwidth will perhaps see a few more shots. The practical problem is that in tournaments Fritz with nullmove searches 15 to 17 ply. 180 seconds x 1.5m nodes a second ==> branching factor is 3.1 in absolute terms (in reality it's smaller but there is an overhead so this model only favours fullwidth search) 4.5^15 nodes or something to get 17 ply = 127236798856 So you need roughly a machine that gets 700M nodes a second to get the same. Apart from the fact that this machine also needs a way better hashtable, because when it is completely overloaded then your branching factor gets worse of course again as you can't profit then much from all the searches that no longer can be stored. Now this model favours fullwidth search still as in reality the branching factor is better for searching with nullmove but qsearch and the first ply give an overhead, whereas in fullwidth you will need even more nodes as soon as you extend a few things. When pruning then extensions matter less for branching factor as the deep nonsense lines get forward pruned quick when nullmoving. So it is a pure theoretical difference. When i let DIEP search fullwidth i do not even get close to 4.5 as branching factor when taking overhead into account. I already need a billion nodes to get to 10 ply then, when all dangerous extensions are turned on... >The interesting question is how many effective plies are lost from null move >pruning. > >Some results suggest that not very much >Tiger14(depth3)-Deep Fritz(selectivity=0,depth 3)35-15(6 draws) >Tiger14(depth4)-Deep Fritz(selectivity=0,depth 4) 31-19(8 draws) >Tiger14(depth 8)-Deep Fritz(selectivity=0,depth 8) 20-23(12 draws) is going to >be finished in an hour. > >The results without null pruning were not significantly better and at depth 3 it >is not clear if selectivity=0 is productive because tiger could win only >34.5-15.5 against selectivity=2. > >Fritz at depth d+1 at the default setting(selectivity=2) seems to be clearly >better than Fritz at depth d(selectivity=0) at small depthes. > >I guess that at big depthes it changes but d=8 is not enough. > >Uri
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.