Author: Don Dailey
Date: 14:41:07 12/28/97
Go up one level in this thread
On December 28, 1997 at 07:24:36, Thorsten Czub wrote: >>* When it is the opponent's turn, select every move >>* Your turn: select the best move (from a SEE+positional 1 ply >>evaluation) and the very threatening moves (including checks of course) >>* Opponent's turn: select several best moves and any threatening move. > >I call this the asymmetrie-search ! >It effects the playing style. >You can try out yourself by using Genius as an email-analyze instrument >(some of my friends play mail-chess) and you will see that genius often >misses GOOD strong moves or sacs other programs see because they don't >prune their >1,3,5,7...plies. >But you can easily force genius to see the sac/overseen moves itself by >taking back one move or playing forward and switch sides. Than the >brute-force part has the job to play with the side that has the >key-moves and will see anything in an instant. > >The senseless thing with genius is that you can let it compute 12 hours >about a position, write down evaluation and main-line and you think you >have a RESULT you can work with. It has maybe computed 9 searches deep >and you think: >Oh - now I am save. >But input the first ply of the anaylzed main line and SCORE AND >main-line will not come back to the same ! Not because of preprocessing >or whatever, but because now the 2,4,6,8 sees different things because >you are one ply shifted. > >If you analyze , after you have sent YOUR move, with genius on the >opponents move something horrible can happen: >You let it again compute 9 searches or 11 and it plays a boring move. > >And than you get a sac by another mail-chess guy who has NOT genius but >e.g. Wchess and the enemy plays a sac and you input the move in genius >and genius sees in an instant that it is lost !! >This effect comes from what I call the asymmetrie-search. >The advantage of this search was on the other hand: >I don't have to compute that many branches. >I am always seing threads. >Sometimes I don't play the killer-move but a 2nd or 3rd best move. > >THIS is the reason genius plays boring ! >If you would see all threads of your opponent, but not your own chances >as good, you would change into a paranoic, careful creature. >This is exactly what Genius is. >Of course Weiner and Lang have tried to change this behaviour over the >years to compete with Fritz and Mchess. But the main problem cannot be >solved without giving up the asymmetrical search from my point of view. >This causes the bori ng >playing style. >Try it out with your own program. But implement it right. >Suddenly your program will develop into what Levy called the: >do nothing but do it right Genius style. > > > >> >>About the "shifted search": usually, the program is careful about what >>could happen to him (looks at a lot of the opponent's moves). But if you >>detect an agressive move (attacking the opponent), it could make sense >>to "shift" and be careful for the opponent (looking at a lot of your >>moves to see if a combination is possible). > > >When I know I have a good main-line > >1 ply 2 ply 3 ply 4 ply 5 ply 6 ply 7 ply >2nd best best 3rd best best 2nd best best 2nd best > > >I could try to prove my 1,3,5,7 moves by shifting the principle. >I know the best defense on ply1 is ply2. But i am NOT sure if ply 1 is >really the best move. > >There must be a method to shift it and to keep the blunders in the >1,3,5,7,9 iterations very small. Genius-level-small. > >One other thing: >I don't think the MxSy indication of the old Mephisto machines has >something to do with plies. > >lets say M2S15. > >Or M4S17. > >Many people always said: >This could mean M2 = 2 plies "brute-force" with overall S15=15 plies >selective peak. > >I don't think this interpretation was ever true. > >Try it out with an old dedicated Mephisto machine. > >Use e.g. the Roma 16 Bit (68000 12 Mhz). >Let it compute 1 second and force move. >Write down the evaluation and the complete main-line and the MxSy = >search-indicators. >And you will be very astonished: >Although M is often 0 or 1 the main line is very good, long, and >accurate ! >HOW can any chess program create these accurate main-lines in 1 second >on such a slow machine without hash-tables ?? > >NO ! >We always misunderstood the MxSy indicator. > >I think I know what it means. But it is very difficult to explain >here (typing it down without examples, high telephone costs, beeing >on-line). >I take my Roma16 Bit to my next tournament. MAybe Paderborn. There I >could show you what I mean by example and we can write down the >main-lines and all this stuff. > >As I said, there is a matrix that correlates the length and the search >depth and (DAN, Vincent, I need your memory-help) when I remember it was > >(2 to the power of search-depth)+1 = length of main-line. > >You can try yourself by taking genius3 and knock off hash and book and >use different selectivity and force different plies search depth as >LEVEL. > >Than you see the matrix. > >From my point of view THIS search was the secret of all richards >programs. >And I am not sure if he can ever come out of this trap. > >>And still another thing: none of the above ideas tells us how to extend >>the search beyond the nominal depth. Everybody here has seen Genius very >>long lines (we are back to the idea of this thread and Junior's long >>lines), going 12 plies beyond the normal exhaustive search. How would >>you do it? > >If you prune heavily in the odd plies, and less in the even ones, and >you have a static-exchange evaluation instead of capture-search, you can >spare much computation time. >The old Mephisto machines indicated search depths components much better >than the new genius. And anything was slower that you could study it >easier. >If it comes to a capture in the main line the root evalaluation of >genius if very very very often bullshit. >All the advantage he had in the 80ties were eaten by opponent programs >using null-move and hash-tables. >Today his concept is rusty. It still works but cannot reach the top >again. >What a pity. Thorsten, Great post! I love reading this kind of stuff. I'm not sure I agree with every little bit of it but a lot of it seems to make sense. The last part I disagree with, that his concept is rusty. I think Richard has not worked very hard on the program for a long time. I think he is capable of comming back and kicking butt if he was inclined to put in the energy. Another problem was that everyone gunned for him. For a while the only program I tested against was his. I figured if I could do ok against Genius I could beat anyone else! That was only a few years ago. But Genius is no dog when it comes to tactics. His secret seemed to be that he found tactics better than anyone but defended better still. I am wondering if the value of his algorithm is related to the power of the hardware? I don't think null move pruning is very effecting on XT hardware but Richards program Psion had long PV's even on XT machines. My first program running on an XT did 3 and 4 ply (FULL WIDTH) searches in tournament time on 4.77 MHZ xt's! I was pretty excited when I got my first 10 MHZ 80286 machine, I thought it to be quite incredible at the time. -- Don
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.