Author: Vincent Diepeveen
Date: 12:54:50 08/18/00
Go up one level in this thread
On August 18, 2000 at 13:50:58, Christophe Theron wrote: >On August 18, 2000 at 07:15:30, Vincent Diepeveen wrote: > >>On August 18, 2000 at 02:00:59, Christophe Theron wrote: >> >>>On August 17, 2000 at 23:16:56, Vincent Diepeveen wrote: >>> >>>>On August 17, 2000 at 22:23:54, Uri Blass wrote: >>>> >>>>>On August 17, 2000 at 21:21:46, Vincent Diepeveen wrote: >>>>> >>>>>>Hello, >>>>>> >>>>>>A few years ago (about 3) i claimed that searching 18-20 ply was possible >>>>>>with huge hashtables, nullmove, a good evaluation function, and >>>>>>several billions of nodes. >>>>> >>>>>searching 18-20 plies with recursive null move pruning is possible and there is >>>>>no doubt about it. >>>>> >>>>>You may miss things because of null move. >>>> >>>>Are you referring here that i need a bunch of plies more for a few >>>>very rare positions where way more as 2 nullmoves are not enough to >>>>see the truth? Just like my program won't find that huge mate as >>>>posted a bunch of messages below? >>>> >>>>>>I was considered nuts by half of the RGCC population, because no >>>>>>branching factor was capable of being that good, when you would >>>>>>search deeper, your branching factor would NOT get under 4.0, that >>>>>>was considered impossible by a lot of people even. >>>>>> >>>>>>Obviously most people following: "i believe that i see in the ads >>>>>>or where i see the outputs from", they challenged me. Some went even >>>>>>that far and called me nuts, a liar and a frog and many other terms. >>>>> >>>>>I did not read RGCC but I believe that people called you a liar because you make >>>>>the impression that you can see everything in the next 18-20 plies except some >>>>>lines that you will look for only 14-16 plies because of null moves when the >>>>>fact is that you can miss also lines with 10 plies because you use recursive >>>>>null move pruning(I remember that this was my impression when I read your posts >>>>>here about 18-20 plies). >>>>> >>>>>I do not say that using recursive null move is wrong but your 18-20 ply search >>>>>can miss 10 plies lines(If your evaluation is good enough these lines are >>>>>usually not important but I believe that there are cases when they are >>>>>important) >>>> >>>>This is incredible hard to believe. The number of positions where >>>>3 free moves + qsearch will fail. If that fails, then there must >>>>be something really wrong in evaluation! It sure can miss things, >>>>but the alternative is to search at most 12 ply fullwidth after >>>>a full night. Give me 18-20 plies then WITH nullmove please! >>>> >>>>So a zugzwang or 2 is no problem to detect then. Apart from that, the >>>>rare times we see a zugzwang arise in a game, that's usually already >>>>seen by my double nullmove. >>>> >>>>With a material only program proof isn't hard btw for this, >>>>as you can take openingsposition where my stupid experiment played >>>>1.a3 searching 30 plies. >>>> >>>>However, the whole discussion 3 years ago, >>>>when i cannot remember any person called Uri Blass posting at that time, >>>>was not a claim of mine that chess could be solved. >>>> >>>>In contradiction!!!!! >>>>Where Jaap v/d Herik writes in the start of the 80s in "computerchess" >>>>the next quote: "when software will search 10-11 ply then no human will >>>>be able to ever beat it". I completely have said the opposite, >>>>that after a ply or 12 only evaluation matters! >>>> >>>>Nowadays i would modify that already to 10 ply, noting that in difficult >>>>openingspositions the understanding of todays chess programs is still >>>>that bad that they need a few ply more to see some consequences there, >>>>so still that 12 there as found by De Groot to be the depth where the >>>>majority of short term plans are based upon is still valid! >>>> >>>>I said this partly based upon also experiences of my draughtsprogram >>>>which already for years can search easily 25-30 ply after a night, and >>>>the only way in which we (marcel monteba and me) could improve the >>>>search was by adding knowledge. This draughts program searches fullwidth >>>>by the way, as doing nothing is usually very well. I don't need to >>>>note that it sees all tactics of world champions within flashes of >>>>seconds, where even an average but smart national player can beat it >>>>pathetically. >>>> >>>>My claim was that branching factor above 10 plies was much better as >>>>i expected it to be, because of better working of hashtables, and >>>>more efficient search by means of a better evaluation which would >>>>basically research the same tree over and over again. Basically i >>>>claimed that with the number of nodes that Deep Blue searched, one >>>>could easily build a much better quality search in software. >>>> >>>>So what you write here above : "searching 18-20 ply is without doubt >>>>possible with nullmove", that's exactly what no one dared to say 3 years >>>>ago, and i'm happy you write it here! It proves how opinions have changed, >>>>as 3 years ago NO ONE dared to say that, except me. >>>> >>>>Now i don't want to sound like a profet, but i wanted to raise this >>>>discussion a bit to show how fullwidth search has superseded by >>>>nullmove driven search, and how these programs have progressed in search >>>>the past years and now are dominating the scene, where 10 years ago >>>>Genius with a fullwidth search, a few singular extensions, pruning and >>>>a clever selective search completely dominated the world of computerchess >>>>with tactics, only losing now and then from faster machines. >>> >>> >>> >>>In my experience, Genius is one of the most complete program in term of >>>knowledge, especially in the endgame. You'll have a hard time if you try to find >>>a position that Genius does not understand when other modern programs understand >>>it. >> >>Genius sees tactical a lot indeed, but it's knowledge is really outdated. >>It's a slaughter to play against Genius nowadays as it's missing too >>much knowledge and it sure isn't good in endgame, the more boring an >>endgame is, the bigger the chance it loses it! > > > >The only explanation I see is that we do not have the same Genius... Or our rating difference is 2255 points, that's also a possible reason. I just missed my IM norm 2 weeks ago. Hope to be IM before 2003. > > Christophe > > > > > > >>This doesn't take away it's great achievements in the past, let's >>be very clear of this. Still at blitz it's a real strong opponent. >> >>>Its search is also very good, that's true, but if Genius did not win against its >>>competition with tactics when it was the best, it would have won with better >>>knowledge anyway. >>> >>>I don't think there are any singular extensions in Genius (at least nothing that >>>goes far beyond the classical extensions that almost everybody does), and I >>>think its pruning system is one of the best ever written by a human being. >> >>Sometimes it's hard to see the difference between smart extensions. >>Can be S.E., can be other extensions as well! >> >>>I also think that its evaluation is still today one of the best. >> >>Here i disagree completely. It's evaluation is completely outdated in >>my opinion, and i see it lose game after game because of it. >> >>>But unfortunately it has big problems somewhere else and nowadays is only >>>competitive on very slow computers... Now I stop, because if I go on I fear Ossi >>>is going to cut and paste my post into his next advertisement campaign. :) >>> Christophe >> >>Oh well, genius has gotten tactical stronger, that's about it. It's >>knowledge is still from 1990 or so. >> >>Hard to make commercial for something like that. >> >>Therefore i'm interested in how Rebel is gonna do in London; whether >>it's knowledge has been worked at, because rebel 9 and 10 i didn't >>find big improvements over 8.And 8 also was i.m.h.o. not a knowledge >>break through, but a speed breakthrough.
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.