Author: Omid David Tabibi
Date: 16:11:59 12/18/02
Go up one level in this thread
On December 18, 2002 at 18:43:26, Bruce Moreland wrote: >On December 18, 2002 at 13:13:09, Robert Hyatt wrote: > >>Actually I found it "non-conclusive" as I reported. It helped in some places, >>hurt in others, and the end-of-term stuff here (and then the SMT stuff on this >>new machine) side-tracked me for a while... I still have plans to play with it >>further. > >So you took an initial stab at repeating this and failed, if I can read between >those lines. > >I implemented this in Gerbil last night and ran it. > >I found that this was inferior to both R=2 and R=3 at every one second interval, >with ECM, between 1 and 20 seconds. > >Meaning, that it never produces more solutions in a given number of seconds. > >General R=3 also is never better than R=2, given this testing methodology, in >Gerbil. > >It is possible that I implemented it wrongly. There are a couple of things that >I don't understand and had to guess about: > >1) When the null-move search comes back with fail high, and verify is true, I >will do a regular search with reduced depth. If this search fails high, I cut >off like normal. If this search does not fail high, I have to re-search with >the original depth. What I don't understand is what I do if the initial reduced >depth search modified alpha. I am assuming that I put alpha back the way it was >and start over. > Some programs have instabilities which result in excessive re-searches using this algorithm. At first stage don't do any re-searches, and when all other things work fine, implement the re-search. The tactical strength is not affected by the re-searches. They merely help avoid zugzwangs. Reducing two plies instead of one, after a fail-high report might work better for some programs. Test them both. Each program has a certain depth, which beyond it, vrfd R=3 *always* constructs a smaller tree. For some programs this threshold might be higher, so conduct your experiments in higher depths. >2) I don't know if this implementation allows two consecutive null moves or >what. In specific, I don't know what "null_ok()" does. I am assuming that if I >don't allow two null moves in a row already, I can continue to not allow two >null moves in a row. > I didn't allow two null moves in a row. >3) I am assuming that if the search cuts off when doing the reduced depth >searches, that the depth record in the hash table should be the original depth, >and not the reduced depth. > I used the reduced depth, as recording the original depth would be misleading. >I can't find any bugs in my implementation, if my assumptions were correct. > Find detailed discussions of all these points in the archives. Omid. >bruce
This page took 0.01 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.