Author: Andrew Williams
Date: 04:33:13 03/23/00
Go up one level in this thread
On March 23, 2000 at 06:48:46, Francis Monkman wrote: >On March 23, 2000 at 06:14:22, Andrew Williams wrote: > >>On March 23, 2000 at 05:16:56, Francis Monkman wrote: >> >>>On March 23, 2000 at 03:09:33, Jason Williamson wrote: >>> >>>>I have been running Amnon 5.06 under Fritz and have noticed really WEIRD >>>>pondering by Amnon. It almost seems that it was just randomly looking at a move >>>>and analysing it, even if said move just dropped a queen in one move, or a piece >>>>or what ever. Mind you, it plays ok on its own time, and still maintains a >>>>position of one of the top 5 or 6 amatuer chess engines I have. >>>> >>>>Jason >>> >>>If I'm not mistaken, this problem is related to the difficulty of extracting a >>>pv line from an MTD(f) search. >>> >>>Francis >> >>This is interesting (to me anyway). My program uses mtd(f) and it extracts >>its PV from the hash table. I found that if the hash table is of a reasonable >>size, the PV for the first several plies is usually OK. >> > >It was just the other day (after some experiments with MTD(f), storing pv in >hash table etc) that I noticed that AnMon's pv line in Winboard was clearly >printing garbage. I note your use of the words 'reasonable' and 'usually' with >interest! > OK. "Reasonable" of course depends on the length of the game to some extent. PostModernist usually uses 256K entries in its main hash table. Sometimes it runs with 512K entries. I only have 64MB RAM on both the machines I use, so my hashtables aren't huge compared to some others. As for "usually", I present these figures with some trepidation: A long time ago I was worried about this issue and I tried to calculate how frequently the score generated by a search corresponds with the score I get if I walk down the PV and evaluate the position at the end. I found that in >90% (I think) of cases with test positions, I got a match. The figure for playing games (against gnuchess I think) was either 59% or 64%, I can't remember which. At the risk of being misunderstood, I think the early part of my PV is "usually" OK. :-) >BTW on this topic, I wrote to Prof Hyatt recently pointing out that I'd found >identicalities in the 17 lsbs (all I'm using for hash address-generation in a >Chess variant, Shatra) of the piece hashkeys' random numbers output by the >'Knuth' algorithm he implements in Crafty. Also that Knuth now (1990s revision) >proposes a fix for said algorithm. I use the "Mersenne-Twister" random number generator, because it sounds really cool. Andrew
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.