Author: Robert Hyatt
Date: 10:37:02 04/24/01
Go up one level in this thread
On April 24, 2001 at 11:25:29, Vincent Diepeveen wrote: >On April 24, 2001 at 09:57:39, Robert Hyatt wrote: > >>On April 24, 2001 at 08:20:57, Vincent Diepeveen wrote: >> >>>On April 24, 2001 at 03:47:15, Uri Blass wrote: >>> >>>>the best software that is not IBM. >>>> >>>>Suppose there is a match of 20 games at tournament time control >>>> >>>>I am interested to know how many people expect 20-0 for IBM >>>>How many people expect 19.5-.5?.... >>> >>>>If IBM expect to do better result then the average result that the public expect >>>>then they can earn something from playing a match of 20 games with Deep Blue. >>>> >>>>I believe that a part of the public who read the claim that kasparov played like >>>>an IM are not going to expect good result for IBM.> >>>>Uri >>> >>>First of all IBM would get out of book every game with -1.0 pawn >>>disadvantage (which is about the average of what Kure and Noomen >>>get in tournaments, sometimes they get out of book with mate in XXX even). >> >>That is pretty funny. I use a very simple to create book, and I don't get >>out of book that badly in _every_ game. Not even every other game. >> >> >> >>> >>>I would expect IBM to lose with 18-2. >>> >>>Let's be realistic >> >> >>Yes, let's. :) 18-2 is pretty funny. >> >> >> >>> >>> a) IBM searched 11-13 ply in 97, nowadays programs search deeper >> >> >>Every time you make that statement I am going to correct it. From the log >>files of the 1997 match we _know_ they searched 15-17 plies deep. Not 11-13. > >First of all the search depth *shows* 11 to 13 ply at most. No it doesn't. Here is yet another log excerpt from an early middlegame position: 5(5)[axb5](60) 60 T=1 pa4b5P Pa6b5p ra2a7 Ra8a7r bc5a7R 6(5)[axb5](57) 57 T=1 pa4b5P Pa6b5p ra2a7 Ra8a7r bc5a7R 7(5) #[axb5](51)##################################### 51 T=2 pa4b5P Pa6b5p ra2a6 Ra8a6r ra1a6R Nd6b7 bc5e3 Rc8a8 8(6) #[axb5](46)##################################### 46 T=7 pa4b5P Pa6b5p ra2a6 Nd6b7 bc5e3 Nb7d6 ra6a7 Bf8e7 9(6) #[axb5](49)#################################### 49 T=55 pa4b5P Pa6b5p ra2a6 Nd6b7 bc5b6 Ra8a6r ra1a6R Nb7d6 10(6) #[axb5](49)##################################### 49 T=160 pa4b5P Pa6b5p ra2a6 Nd6b7 bc5f8B Qe8f8b ng3f5 Ra8a6r ra1a6R Rc8b8 11(6) #[axb5](49)#[Nf5](50) 50 T=308 ng3f5 Nd6f5n pe4f5N Pb5a4p bc2a4P Bd7a4b ra2a4B Qe8d7 bc5f8B Rc8f8b pf5f6 Qd7d5p qf2g3 Pg7g5 --------------------------------------- --> 33. Nf5 <-- 7/65:41 --------------------------------------- Again I reiterate, the notation 11(6) means 11 plies in software search, 6 plies in the hardware, plus the quiescence in hardware. There is _no_ argument with this. Simply ask any of the DB guys. 11(6) is a total of 17 plies of search. > >Secondly it is theoretic impossible to search 19 ply fullwidth (13 ply >software + 6 ply hardware) without hashtable last 6 plies. no it isn't. And I will be glad to show you a 19 ply search by Crafty if you want. Might take a while, but it is _not_ "theoretically impossible". It is theoretically possible to solve the game, in fact... > >This is simple math. > >how bigtime it suffered from not using hashtable can be easily >seen in game 2 for example. You need to try the experiment first. I did so a few years back when having a discussion with Hsu. It doesn't hurt nearly as much as you think. Even Junior doesn't store the last ply or two before q-search. It hurts less in the middlegame, more in the endgame, but not by a huge amount... the test is easy to do (for me anyway). > >One of the moves Be4 which it made where Kasparov shortly was of >the opinion that deep blue team 'cheated'. There there are many >transpositions possible. Despite that deep blue doesn't search >deeper as in openings positions where there are little >transpositions possible. Your math isn't working here. The effective branching factor at the Be4 position is _still_ quite high. It is not a simple endgame yet. > >I think Kasparov later corrected that, but i'm not sure of it. > >In diep the average number of moves is 40 in middlegame on average >(endgame of course not counted). That is because it sees >many stupid nonsense moves which humans do not consider soon. > >With hashtable you would reduce that bigtime (or as you indicate >reduce actually seach depth needed, whatever you do, you will get >a better b.f. as result). > >Last of all i DID do experiment with DIEP searching fullwidth and not >using last 6 ply hashtable. > >Could you do this with crafty too? Of course after you also added SE >to it for the first so many ply minus 6 and also turn on more recapture >extensions and turn on checks in qsearch to some extend (for example >only first ply). > >Now we can compare. Please posts nodes and search depth needed. > >The good b.f. which DB seemingly has first few seconds >is a result of that initially it can't put 480 chessprocessors >to work very efficiently within a few seconds. Sorry, but DB used 480 chess processors for _every_ search it did. That was the way the algorithm was written. As a result, its search (and branching factor) for early plies was worse than for deeper plies. This all explained by Murray several times. > >Nevertheless, please do the experiment and report back. For diep i need >billions of nodes for 10 ply already! I'll try to find my old data on this. But if you need billions of nodes, something is broken. I can turn hashing _off_ and not need billions of nodes to search 10 plies... > >And i'm pretty sure i sort my moves better as Deep Blue ever did! >Not to mention my evaluation is way better! I love statements of pure subjective opinion when given as uncontestable fact... Since you have never seen their evaluation in particular... > >All those factors are completely irrelevant of course. The proof for >their search depth is so obvious! > >Apart from that studying logfiles you see that they get fail high to >some tactical moves at very explainable search depths. > >Like the tactical move Bf5 in game 6 is 8 ply for them. With recaptures >and SE i also need 8 ply for that. In fact most programs which by default >do recaptures already need 8 ply! > >>15-17 in the middlegame, more in the endgame. I don't know where you get the >>11-13 nor why you keep saying it when the log files clearly show that is wrong. > >6 ply in hardware + 13 ply in software = 19 ply fullwidth. >Not 15-17. 11-13 ply is what they got. If you would add 6 ply in hardware >to that that's 17-19 ply fullwidth! That is _exactly_ what they got. 11(6) _does_ mean 17 plies + q-search. I am not certain that the last 6 plies was 100% fullwidth... that I don't know and I haven't taken the time to ask them for real details. They did mention futility pruning in the hardware, so perhaps some of the last 6 plies were subject to this. Or perhaps just the q-search was using futility. But there are numbers in the log that match what they have said. I don't see any way to dismiss (a) log files and (b) direct statements from the team members. > >We only get that with nullmove AND hashtables! > >Please do the experiment Bob and as only programmer defending >Deep Blue here it's very obvious that the facts are hard to ignore! It seems to me that you are doing a pretty good job of "ignoring the facts" here. Facts given directly in the log files for example concerning their search depth. I often reach depths of 14-15 in middlegame positions using null-move. Turn it off and that drops by 2-3. But then make me 200 times faster and I gain that back and _then some_. > >> >> >>> b) their book is hell worse as nowadays books are >>> c) positionally it never was good, it doesn't even >>> know what a good bishop is nor when a doubled pawn is >>> good (f2,g2,g3 pattern happened twice in games against kasparov) >>> also it exchanges sometimes queens in a position where not exchanging >>> wins for IBM >> >>Let me know when you or your program beats Kasparov in a standard game. Or >>even when you _draw_ him. Then tell me how weak they played positionally. >> >> >> >>> d) hardly can use EGTBs >> >> >>"Hardly"??? used them just like I do. They used them _before_ I did in >>fact. They were using them in the late 1980's. Just like HiTech did. And >>others. >> >> >> >> >>> >>>So in *all* respects it is getting outgunned. Not to mention EGTBs. >>> >>>No one talked about that subject yet, but last so many plies they can't >>>use EGTBs. They only can use them the first 5 or 6 ply, that's it. >> >> >>They use them in the first 11-13 plies. Which is _exactly_ how I use them. >>I don't probe in the q-search. I don't probe beyond the basic nominal search >>depth. It works fine for me. It works fine for them. They don't do it in >>the hardware part of the search, which means the last 4-6 plies plus q-search. >>That is _not_ a problem since it works well for me. >> >> >> >> >> >>> >>>Though this is very good compared to not using them, this means simply >>>that all exchanges towards a lost 5 men they will not detect. >> >>And if you probe in the q-search you will get killed tactically when there >>are only 6-10 pieces on the board. You will lose _several_ plies. I know. >>I have been there. >> >> >> >> >>> >>>They lose with induction to everything. The level of software has increased >>>bigtime when compared to 1997. Of course the strategical problems are >>>still there and some positional problems are still there, but in >>>computer-computer games you hardly can take advantage of that. Only >>>a human versus a computer can! >>> >>>Best regards, >>>Vincent >> >> >> >>Your "induction" is broken. Make that "non-existent". HiTech or Cray Blitz >>won't lose to "everything" today by any wild stretch. Much less Deep Blue. >> >>The jealousy directed at this program/project by chess programmers (some anyway) >>is remarkable... the speculation about how it operates is even more remarkable. >>And finally the amount of disinformation about it is unbelievable.
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.