Author: Robert Hyatt
Date: 05:39:08 10/24/00
Go up one level in this thread
On October 24, 2000 at 02:10:57, Christophe Theron wrote: >On October 23, 2000 at 22:48:12, Robert Hyatt wrote: > >>On October 23, 2000 at 20:50:35, Ratko V Tomic wrote: >> >>>>>> GT has larger-than-life king safety scores. That >>>>>> is all. No different search paradigm or anything else. >>>>... >>>> Christophe specifically said that the search is the same for both >>>> programs. I took that as being true,as it seemed to match my >>>> impression after watching games. It speculates more. But it >>>> isn't searching _differently_ at all. >>> >>> >>>There is quite a bit of space in between, to paraphrase your earlier comment, >>>'large safety score and nothing else' and your current one 'is not searching >>>differently'. The hypothesis I sketched in the previous note is just one of many >>>conceivable ways in between, where one can say 'the search is the same' (i.e. it >>>is using the same iterative alpha-beta & support routines), yet the novelty >>>cannot be said to be merely in a new leaf evaluator (or its weight) but it would >>>be in a way how the inputs and outputs of the search are interacting across >>>iterations via pre/post-processors, as well as how much information is >>>transferred that way (is it just score, hash, history entries & killer moves, as >>>in most programs, or something extra which helps GT make fewer mistakes in >>>deciding to undertake apparently open ended king-side attack). >>> >>>While Christophe did say the search is the same (and one can parse that to mean >>>many things; even Botvinnik's program had alpha-beta search in the lower layer), >>>he also suggested, in response to dismissive comments about the GT style as >>>being just another king safety tweak, that you're welcome to go ahead and >>>increase the king-safety scores in Crafty and see how far that gets it. >> >> >>How do you think I arrived at the _present_ king safety scores? Here is the >>point: >> >>if you have a good search (and CT certainly appears to meet that criterion) >>so that you don't get out-searched very often, then you can be more speculative. >>If you do get out-searched, then you will have massive problems, as CSTal did >>in every group of games I watched it play vs Crafty. >> >>I'd be willing to bet that I can tune my aggressiveness way up, _and_ play that >>version using a big alpha machine (to be sure I don't get out-searched anywhere >>along the way) and the aggressive version would do fine. But as hardware >>becomes more equal, then the 'speculation' had better be right. Else the >>more accurate search will find the holes in the speculation and blow through >>them. > > > > >In many cases we are talking about refutations that can only be found with a >20+ plies search (see 43.Rc6 in the Gambit Tiger - Nimzo 8 game). This is above >the current computers/programs abilities anyway (including mine). > >So it works also when my program is outsearched (which indeed does not happen >often). > > > > Christophe > > > One position doesn't mean "it works". You point to Rc6, which turned out quite good. I saw a game it played vs a GM on ICC where the speculation did not work, it ran into an attack it didn't understand, and got zapped, royally. I also noticed that normal tiger seemed to do as well or better against programs in the tournament someone posted partial results for here... > > >>We actually played tuned like this in several ACM events using Cray Blitz. And >>it worked quite well since we were out-searching all the micros by huge margins. >>But against more equal opponents like deep thought and hitech (and belle in the >>early 80's) this was not a wise thing to try.
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.