Author: Andrew Dados
Date: 11:48:03 09/17/99
Go up one level in this thread
On September 17, 1999 at 12:26:51, Owen Lyne wrote: >Ah - but Bob rejects that due to what happens when you are wring. You've perhaps >serahced some useless move (and wasted time/cycles doing it) before using >some history/killer or some such good move, which would have cut off and saved >all that effort. So unless you really believe the move is good, it is perhaps >not worth even the effort of picking the 'best' out. > >I do hope I got that straight... ;) Read Bob's posts in the big hash table/PV >discussions. > >Owen Bob did experimenting with his 'lazy eval' being too lazy and totally anti-fail-soft. What it did - it returned alpha on absolutely worse moves (like losing a piece), which denies purpose of fail soft - that move would get to hashtable and screw branching factor instead of improving it. If, however, BF would improve with lazy eval turned off (or returning something other then alpha maybe), then only question is - at which ply fail-soft improvement of BF offsets speedup of lazy eval (and with what HT sizes). For each program/qsearch implementation the answer can be *totally* different: most likely the lower your flip rate for 'all' nodes from ply to ply (only those 'flipping nodes' are of our interest here) - the less improvement in move ordering expected... ymmv, as usual :) One more loose remark here: while our node on previous search failed low (means all moves were searched and returned < alpha) it may be interesting to try *all* captures at the end - most likely no easy capture will fail high now... or at least do killers/history moves before any captures. Which can be an interesting thing to experiment with... -Andrew- P.S. below is the line from crafty lazy eval code: if (tscore+largest_positional_score<= alpha) return(alpha); To improve fail-soft scores with lazy eval, one may try changing that to something like return(tscore+largest_positional_score).
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.