Author: Robert Hyatt
Date: 12:29:20 11/09/00
Go up one level in this thread
On November 09, 2000 at 13:52:44, Bruce Moreland wrote: >On November 09, 2000 at 03:22:19, Joe Besogn wrote: > >>There are some posts under I made which endeavour to explain it. In the thread >>"for those interested in new paradigm", or whatever it was called. >> >>I found your answers difficult because of mindset and word-meaning >>discrepancies, but I think what you're saying is that null-move was a >>revolutionary change because later "we kicked ass". Meaning Lang and presumably >>Schroeder programs. > >It was a dramatic increase in strength for those who wanted it. I doubt that it >affected the top level because they either had it or had ways to deal with it. > >>In Kuhn terms, the crisis state you describe was that 'amateur' programs were >>weak. >> >>Hyatt's response is that null-move was a typical field evolutionary development >>over 15 years, from Kaissa, Berliner, through Beal and Donninger. > >When I first met Bob, he did not sound like a null move believer. In my first >conversations with him he didn't sound like someone who thought it would >increase strength massively, and I am pretty sure that when he first did it in >Crafty, he used R=1 rather than R=2, and may not have allowed multiple >null-moves in the same path. Actually when I first started using it in Crafty, I started off with recursive R=2. But that was in the days prior to even the P5/133, which meant 5-6 ply searches in blitz were common. And there it wrecked total hell with search accuracy, hiding mate threats that a beginner could spot. I changed back and forth between R=1 and R=2 for at least a year, until the hardware finally reached a point where 5-6 ply searches no longer happened. Then the failures became less frequent. On ICC, in blitz, I now typically search 10-11-12 plies deep. And the failures are almost non-existant. > >If null move is in Kaissa, it is not in the literature that I have seen. Null >move in the Berliner program was published in 1990, unenthusiastically. I first heard about the "idea" in 1977 at the WCCC in Toronto. Donskoy didn't say a lot about what they were doing, but he overheard me talking to someone (I think Ken Thompson) about my threat detection using this. Note that we didn't call it "null-move" back then. I called it an "any" move in one context and a "pass" move in the other. The advantage for me back then was to make a move, then pass, then find out what good things I could accomplish after the pass. That meant that the move I just made prior to the pass exposed new threats (since I had already checked for good things before the pass as well) so I knew what threats existed already. Donskoy tried to explain how he used the "null" move, but things were hectic at the WCCC, he was from the Soviet Union and so didn't speak great English, and as result, we never got very far into the inner workings. But several have mentioned that he certainly was doing _something_ with them. Patsoc (the program developed in Berliner's thesis) was a 1970's program as well, and it used the null-move idea in various forms as well. I don't think anybody used it in today's form, back then. But by 1981 or so I was using null-move R=1 (non-recursive) after being given something written by Don Beal that described this. (I could probably thumb thru the old Cray Blitz source to possibly discover when it was added, although I don't think I kept a journal of changes in the comments like I do today in Crafty). > >I think there might be two Don Beal null-move articles. The one in ACC5 (1989) >is mostly about null-move quiescence search. I wouldn't read that article and >rush out and add null-move forward pruning to my search. I would read it and >realize that the Belle quiescent search model had null-move built in. > >These guys are publishing about the cardiac effects of nitroglycerine. They >neglect to mention that you can blow the hell out of things with it. That is of >course much more interesting and fun, which anyone who has taken nitroglycerine >would agree. > >It was the Donninger article that convinced me that null-move was not just some >goofy academic thing like "chunking" and so on. > >>Null-move is forward-pruning, right? Lang and Schroeder did this forward pruning >>at that time by working on some chess knowledge at each node. Null-move does the >>forward pruning by a depth-reduced search at each node. > >I don't think it was enthuiastically adopted by the amateur/academic set until >the Donninger article. That's most of the programs, and it's almost all of the >people who will talk aout programs. > I can't answer for the amateur set, but it was used in the early 80's in more than one academic program. If you can date Don't first paper, I probably was using it by the time that was published, as I believe that he sent me a preprint to read before it was published. I want to say it was "selective search without tears" but I am not sure that was the first one. >>Schroeder, I have read, doesn't use null-move, even today. >> >>So, if you claim that 'amateur programs' were in one paradigm before nullmove, >>and then another after, you are going to be implying that Schroeder/Lang >>programs are in yet another one, now, and then. Unless you argue that the two >>forward pruning techniques (knowledge vs search) are sufficiently conceptually >>identical that the 'amateur' paradigm merged with the Lang/Schroeder. Too messy, >>I think. >> >>Hyatt's model of nullmove as evolution seems stronger. imo. > >Maybe. It's not like I really know what I'm talking about. I just think that >the enthusiastic adoption of null-move a) created an immediate major increase in >strength amongst the non-professionals, and b) was triggered by the Donninger >article, regardless of what articles had been written previously. > >bruce One important question: how much does R=2 recursive null move improve a program. I haven't tried this in a good while, but I believe that the last time I ran a huge match, one with, one without, the result was maybe a difference that could be expressed via Elo's formula as about 50 points. That really isn't a wild jump. IE I would bet that the basic idea of check extensions might produce more than that, although I haven't tested it that I remember.
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.