Author: Vincent Diepeveen
Date: 02:00:27 06/28/98
Go up one level in this thread
On June 27, 1998 at 13:43:50, Don Dailey wrote: >On June 27, 1998 at 10:15:40, Vincent Diepeveen wrote: > >>Hello Bob you're lately telling us a lot of new facts, >> >>I remember that some time ago (2 years perhaps?) >>we already knew the DB speed. i claimed that getting 18-20 ply >>would be no problem for them if they stripped some extensions, >>improved move ordering and used nullmove. >> >>At that time you told your audience that getting 20 ply was >>*impossible* and challenged me. >> >>Right now my program GETS 20 ply in several middlegame >>positions after 24 hours of analyses (and i'm not referring to >>positions where you need to make a forced move like recapture), >>but i'm also in the pruning buiseness, so probably this doesn't count >>in your eyes as some stupid variations get pruned by Diep. >> >>Note that very important is the fact that i'm having 80MB of ram now, >>under NT i use 60MB of hash under DOS 64MB. It appeared that >>big hashtables at analysis level give a lot (which is rather unsurprising). >> >>Now how big was my surprise when reading next: >> >>>Again... if DB used null-move R=2, with a search like mine, they would do 20+ >>>plies in middlegame. They do extensions instead. I do 12 plies at 200K nodes >> >>How can they at hardware which is slower than was expected few years >>ago (i still remember an email from one of the teammembers very well >>where they expected to make a chip which got 5-10M nodes a second, >>and that has become finally 2.5M) >>get suddenly 20+ ply without my 'dubious' forward pruning, but with >>the stuff discusses in RGCC, and they even could search deeper >>in your opinion? >> >>This needs some explanation! >> >>>per second. They are 1,000 times faster... with a branching factor of roughly >>>2.5, what is log base 2.5 of 1000? That's my search depth at that speed. So >>>*obviously* they are doing something else. You compare your 12 plies to their >>>12 plies. I say that's total hogwash. >> >>Ok ask IBM for old printouts of the match and we can easily compare their >>mainlines to our mainlines. >> >>Don't do whether their 12 ply is in fact 24 ply or 13 ply. >> >>12 ply is 12 ply, if you use nullmove you run zugzwang risks. If you don't >>do checks in q-search you even miss obvious mate threats last couple >>of ply, if you prune a lot you might sometimes miss something like >>heavy sacraficing for a mating attack; but lucky none of this all happened >>in the games. >> >>Their 12 ply aren't more holy than mine 12 ply. I admit: i prune, so certain >>stupid lines which might win might get pruned. There were no >>difficult ! moves played by Deep Blue, no difficult sacrafices, no >>mate in 40 announced, and win in 23 ply (23 ply for Diep: Kf1? in game >>II, where Kh1! wins forced) was missed by Deep Blue. >> >>Yet it plays a horrible move which is only 1 ply deep (12.Bxg6? in game 4 >>after the moves: >> 1. e4 c6 2. d4 d6 3. Nf3 Nf6 4. Nc3 Bg4 >> 5. h3 Bh5 6. Bd3 e6 7. Qe2 d5 8. Bg5 Be7 >> 9. e5 Nfd7 10. Bxe7 Qxe7 11. g4 Bg6 12. Bxg6 hxg6) >> >>This move Bxg6 can be easily prevented if a program knows that >>after hxg6 the simple pattern g6,g7,f7 open h-file is not a bad doubled pawn. >> >>This is a beginners fault of 1 ply. So how deep did Deep Blue search? >> >>Please test your programs at it. If i remember well this pattern is also in >>psion. >> >>Vincent > >Hi Vincent, > >I'm not sure comparing principal variation length is a good way >to compare. Probably best is just looking at a lot of tactics >to give a true sense of what is being seen and how soon. But >even tactics are misleading as I can get tremendous improvements >in tactics while weakening the program. I think it's becoming >more and more important to see quiet moves since computers are >rarely going wrong in tactics. Thanks, that's what i think too. But no doubt both DB and Diep see the same, so the only diff is that DB might seen is some more tactics, which was not there in the games. So if they searched 11 ply, and mine usually gets that at tournament level (only few positions i don't get 11/12 ply), then i think positional depths are definitely comparable. What Bob claims is that they saw more because they search more crap. I don't believe that they saw effectively more than other programs. >But what kind of test can we give to show what a computer is >really seeing? This is a difficult problem. We must also >know what it is deducing from positional heuristics. I think >the right kind of tactical test would work the best although >nothing will be perfect. I would want to see a tactical set >that was not overly sensitive to checks, mate threats and >captures, usually extension algorithms that deal with these >moves improve tactics much more that actual chess play. Actually we should consider a lot of 'development' moves right after opening too. Like: most programs want to quickly castle short, instead of delaying castling, other programs prefer initially to always capture that bad bishop of the opponent or bishop which gives the opponent a lead in tempi. Positions like such, simple positions, where you may not make a single error. Good start are the DB games. DB made many faults, neatly analyzed by GM Seirawan. Everywhere where there is a ? or ?! mark of Seirawan we should analyze with out programs. In fact this is what i'm already doing. It turns out that so far i haven't found a program which makes positionally weaker moves than DB did. >Based on discussions I've had with Shu and Murray, my opinion >is that they do way too many of these extensions. They have >told me they can lose 2 or 3 ply and seem to have every possible >extension algorithm known to man. What Hsu/Murray say is of course not true there is too much room for inventing new extensions. It seems that when you search 11/12 ply with 200M nodes a second that you definitely do something wrong to me. So the room for inventing new extensions gets limited: you need to invent extensions that don't hurt your branching factor. > This means anything that >anything they need to see that slips through their network of >extensions can take over a hundred times longer to see! A >lot of ideas in chess are much more than check-check-check-mate >but are more sophisticated things that require seeing subtle >maneuvers and such. > >The rule of thumb I use is that it is significantly more important >to "minimize the worst case behavior." Making it sees a flashy >tactical shot 10 times faster is probably bad if the price you >pay is that all other things are 2% slower. Well i think you should extend some tactical stuff that otherwise would lead to tactics which is usually too deep to see. I think you should do some checks and mate extension, but generally those extensions are cheap to do (and therefore everyone does). I think you need to do mating extension because generally a mate is way longer than the depth you search. >In almost all sports and human endeavors the principle is the >same. The saying "a chain is as strong as its weakest link" >makes a lot of sense. I fully agree. >Selectivity is exactly the same double edged sword. The more >of it you use, the more vulnerable you are to shallow tactics >but also the more deep tactics you pick up in a given amount >of time. It is very easy to go overboard and suddenly you >have a chain with a VERY weak link. Selectivity like forward pruning is very dubious, i prefer to do without forward pruning, but i fear i need to in order to get that 12 ply. Also what i see is that if i know for *SURE* that i get this 12 ply, that selectivity isn't that bad. There are 2 tricks which are in test sets cq. known to the audience which takes Diep now 18-19 ply instead of 14 ply to find. I think that kind of cheap price to search average 2-3 ply deeper. So i doubt whether limited selectivity is that bad, because at the same depth i'm still seeing the same except for some very very weird unforced sacraficial stuff in order to mate the opponent. >The main thing I've noticed about all this tactical stuff is >that you never really win a game on tactics. The win had to >already be in the position. Sometimes you will play a stupid >move that you would not have played if you saw tactics better, >but most programs are not really focused on this problem, most >test sets are designed to see if your program can find a win >in a won position. I have yet to see Deep blue or any other >program find a win in a lost position. I completely agree with this. You first need to have that excellent position in order to win. Yet the many blitz games played at internet lead for many people to conclude the wrong thing, as simply the tactical depth there is around 7 ply for most programs in blitz. So i like to add to this that you need a certain 'human' depth (around 6 moves), in order to make tactics not your weakest link. When that condition gets fullfilled then i prefer the positional better program, so i definitely not pick DB processor then :) Still this internet play is 'troubling' peoples mind. >- Don Vincent
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.