HOMEJOIN ICCHELPMEMBERS & ACTIVITIESCHESS.FMRESOURCESSTORE HOME JOIN ICC HELP MEMBERS & ACTIVITIES CHESS.FM RESOURCES STORE
menu

Right this minute there are 727 members online, of whom 39 are grandmasters and other titled players.


up HOME

"ICC is the most humorous and enjoyable club I've ever visited."
— Grandmaster Michael Rohde


Get a seven day FREE trial of ICC!

ICC Interview 2 with Feng-Hsiung Hsu (CrazyBird)

SJLIM: Hello all.

CrazyBird: hi.

CrazyBird: sorry for the time screwup, missed the email from sjlim.

SJLIM: just a few moments to get more listeners =)

CrazyBird: mostly programmers here?

SJLIM: folks.. we'll be giving priority to the tech questions first.. if we exhaust them.. then we'll carry on with general questions - time permitting.

CrazyBird: this is intended to be more technical, but i guess anything goes.

SJLIM: May I formally welcome Deep Blue creator Feng-Hsiung Hsu aka CrazyBird.

SJLIM: I guess we can begin?

CrazyBird: sure.

SJLIM: here is a long one..

SJLIM: Hello, Dr. Hsu. My question has to do with the (sadly unlikely) possibility of your undertaking a future chess project using DB-like chess chips.

SJLIM: We know that you have acquired the rights to the DB chip design from IBM. It has been reported that in one recent talk you gave you stated that IBM had retained all rights to DB's evaluation function. So, really, two questions:

SJLIM: (1) Just the evaluation function? Or most or all of the final DB software?

SJLIM: (2) How can a new team effectively recreate DB's eval function (or more) without you, consciously or not, impinging on IBM's intellectual property?

CrazyBird: i only have the right to the chip design. also, i cannot reveal what is not already publicly available.

CrazyBird: i don't have the code to the software. I wrote the initial code though, so i can replicate the search code at least.

CrazyBird: theoretically, you could try to license the IP from IBM, but it would be hard to make sense out of the schematics and so on.

SJLIM: thank you.. next question.

SJLIM: What proof can you offer that Deep Blue 1997 was stronger than Deep Fritz 2002? There is little in the six games on record, and the result against Kasparov was not more impressive (chess-wise) than Fritz's against a very well-prepared Kramnik.

CrazyBird: this is all based on old data. deep blue chip was at least 200 points better than the top commercial programs at comparable speed, and deep blue was 100 times faster than deep fritz.

CrazyBird: it was both tactically stronger and positionally better.

CrazyBird: the tactics apparently did not matter in kramnik match. kramnik was not playing very deep tactics.

CrazyBird: the positional part matters in two games, but then they were compensated by misplays on kramnik's part.

CrazyBird: maybe deep blue overshoots in tactics, or maybe kasparov just played better.

CrazyBird: we will know for sure when kasparov plays deep junior. deep fritz is obviously not stronger than deep junior.

SJLIM: interesting...

SJLIM: next question.

SJLIM: What do you think of Brutus, ChessBase's FPGA hardware chess system currently under development? ( http://www.chessbase.com/newsdetail.asp?newsid=221 )

CrazyBird: very interesting. i am doing something similar with shogi. i may come back to chess when that is done, just for the hack of it.

CrazyBird: of course, they have the disadvantage of not knowing what was necessary to play positional chess at very high level.

CrazyBird: but they certainly bear watching.

SJLIM: Welcome OldPhoenix.. message me your questions.

SJLIM: What are the chances of Hsu writing a book about DB targeting the more technically oriented, such as programmers and perhaps the hardware guys? Wouldn't such a book represent an invaluable contribution to the field?

CrazyBird: well, there are two articles out there on deep blue. one in IEEE computer, and one in AI. we pretty said whatever needed to be said.

SJLIM: thank you.

SJLIM: real techie one now.

SJLIM: Question for CrazyBird: In the AI article, you estimate that DB searched an average of 126M nodes/sec. Is that taking into account parallel overhead, or would the equivalent number of serial nodes be much less?

CrazyBird: actually that is Joe's estimate. maybe it is some recent measurement. during the match, it was about 200 million.

CrazyBird: the 200 million was raw count. i don't know what joe meant by the new number.

SJLIM: thank you.

SJLIM: Unfortunately, we have no super machine for computer chess anymore, where a few years ago there were several: Deep Blue, Cray Blitz, etc. How big are the chances that you sit together with Bob Hyatt at the University of Alabama, you do the harware part, Bob the software part and we get an new chess monster?

SJLIM: Since a sponsor would be needed the outcome might be Deep Coca Cola?

CrazyBird: well, building a deep blue class machine is not that big a deal these days.

CrazyBird: in a talk i gave two days ago, i estimated that the bill of material for a deep blue class machine is less than Kasparov's one-day appearance fee...

SJLIM: heh =)

CrazyBird: actually probably less than deep fritz's hardware cost.

SJLIM: amazing.

SJLIM: Here are a couple of yes/no type questions.

SJLIM: 1.Does the 12 plies brute force depth of deeper blue means no pruning or can it include pruning in the hardware?

CrazyBird: it does include some hardware pruning at the last 3 plies. the pruning appears to have no effect on the search result. that is, it is effectively pure speedup.

SJLIM: another,,

SJLIM: 2.Does 12 means that the depth of the software in deeper blue was less than 12 plies(12-x when x is the depth of the hardware that is not constant)?

CrazyBird: yes, the software "brute force" depth is always less.

SJLIM: Question for CrazyBird: In your previous chat session you said that it was possible to "solve" chess.. it was also said that there might be 10^40 positions in chess. This number is so huge, wouldn't it be theoretical impossible just to find a storage media for that kind of data?

CrazyBird: well, i meant it might be possible, not i think it is possible.

CrazyBird: yes, 10^40 is a very large number, but to have the solution tree, you don't necessarily need the full set.

CrazyBird: still, i don't really believe it will be solved.

SJLIM: thank you.. here is a follow up question.

SJLIM: Question for CrazyBird: Before, the topic came up about computers "solving" chess. To some extent, they are able to do this with tablebases. How many tablebase-men do you think can be solved within the next few years, and further down the line? All positions with "x" men or less?

CrazyBird: the number goes up exponentially for each additional man. hardware speed and storage density doubles every 1.5-2 years, but we might be reaching some limit soon. I think we can one additional man within the next 5 years.

CrazyBird: also some the 6-men are not really 6-men, but constrained 6-men.

SJLIM: so that would be 7 man endgame tables then.

CrazyBird: in the sense, pawns are locked and so on.

SJLIM: I see.

SJLIM: next question..

SJLIM: Yace-Author : Can you please clarify, what you mean by "comparable speed" in "deep blue chip was at least 200 points better than the top commercial programs at comparable speed"

CrazyBird: or locked pawn situation, we can go up a little bit faster, since they have a smaller multiplier.

SJLIM: opps.

CrazyBird: no problem. that was the last part of my answer.

SJLIM: ok.. =)

CrazyBird: back to the next question.

CrazyBird: we had a "phantom queen" problem for the 1997 version of deep blue chip.

CrazyBird: which forced us to effectively slow it down by about a factor of 10, and became roughly the same speed as the commercial programs of the day.

SJLIM: I see.. so when they played at this 'speed' you still find a 200 point advantage to DB crippled?

CrazyBird: murray played 10 games with it against the top programs then and beat them 10-0, which gives a reasonable certainty that it was stronger by at least 200 points.

SJLIM: I see.

CrazyBird: the games were very intriguing, because we were seeing repeatedly some hardware evaluation features at play.

SJLIM: interesting.. which ones? =)

CrazyBird: murray is giving all the games he has to icga, but i don't know whether the 10 games are included or not.

CrazyBird: i can give some examples.

CrazyBird: in one game, the opposing program just have no idea that despite its material advantage, its king was getting killed.

CrazyBird: in another, the other program did not realize that bishop of opposite color ending was lost for it.

CrazyBird: or they had no idea that the open file that their rook occupied was just useless.

CrazyBird: something like that.

SJLIM: thank you.

SJLIM: Alot of programmers on CCC have asked me to ask you this.. for clarification..

SJLIM: Please explain search depths for the notations 4(5) and clarify earlier comments about 12(6). This may include indicating what is "normal full width" searching, extensions, quiesence search, or other types of searching DB2 utilized, and which was done in software versus in the hardware chess chips.

SJLIM: Also, what types of pruning were used. This topic has generated enourmous discussion on CCC.

CrazyBird: 4(5)means the same thing. 5-ply maximum hardware depth, although it is obviously impossible in this case.

CrazyBird: since the brute force depth is 4.

CrazyBird: i can't really go into the details of the hardware pruning. it is related to method of analogy pruning, or rather a basterized form of it.

CrazyBird: limitation in the contract with ibm.

SJLIM: Can this be answered? - Does 12(6) mean the 6 is included _in_ the 12, or in addition to the 12?

CrazyBird: 6 is part of 12, but the hardware can search less than 6, that is the software horizon may be more than 6 plies.

CrazyBird: and of course, the selective depth can be arbitrarily deep, well, no more than 8 times brute force.

SJLIM: please "message SJLIM" you questions folks.

CrazyBird: i am curious, anyone received the book yet? the local bookstore does not have it yet.

SJLIM: some people have quoted from your book on CCC I believe.

CrazyBird: argh, the q search. it is in hardware. both sides are allowed checks in quiescence search. max is 8, i think.

SJLIM: thank you.

SJLIM: I think we have only one more tech question for now..

SJLIM: I have a question - it's about Game 6 in the 1996 match. Did DB think that 20 Bxh7+ was a draw? And if so, what does CB think about Berliner's analysis showing that this move would win?

CrazyBird: game 6? Kasparov was winning all the time. are u sure that was the game?

SJLIM: hmmm.. anyone know? =) I guess its the game with Bxh7 .. game 6?

SJLIM: guest211(U) tells you: Kasparov played 20. a3 there, but there was a lot of talk about Bxh7 being a tactical win.

CrazyBird: that may be the case, but what is the point? he was winning already.

CrazyBird: i think berliner was referring to a different game?\

SJLIM: hmmm.. I guess we'll leave that for future analysis perhaps. =)

CrazyBird: there was another game that he could sac on h7, but elected not to, and was glad he did not when he saw deep blue's reply in our lab.

SJLIM: one last comment/question from the programmers..

SJLIM: Here's a question. CB, I appreciate your willingness to engage in this Q&A, but its value is limited due to brevity and lack of followup.

SJLIM: Would you consider joining a moderated computer chess message board, such as the Computer Chess Club, in order to develop a more robust and full discussion of the many questions surrounding the programming and performance of Deep Blue? I am certain we programmers would welcome your participation.

SJLIM: And, it's always possible that your participation in a public forum might encourage potential sponsors to work with you.

SJLIM: by the way Hsu, here is a message from Jack who is joining us in the discussion now..

SJLIM: Jack (13:55 19-Oct-02 EDT): I received book from Amazon yesterday

SJLIM: =)

CrazyBird: well, i am retired as far as computer chess is concerned. besides, it is not clear that there is a great demand for something like deep blue to come back.

CrazyBird: and being a married man means priority changes:)

SJLIM: no doubt =)

SJLIM: ok.. we are running low on time.. I'll try to sneak in as many questions as possible =)

SJLIM: ophir : it was described by Mr. M. Campbel that DB lost game 1 in 1997 becuase of a "random move" - what does that really mean?

CrazyBird: argh. it was lost to begin with. a bug terminates the game early and caused the kasparov camp to spent all night analyzing why.

CrazyBird: they reached the conclusion that it saw a very deep mate:).

SJLIM: thats hilarious =)

SJLIM: what kind of bug?

CrazyBird: it was something related to move selection, some data structure problem, i think.

SJLIM: I see.. moving along.. quickly.

CrazyBird: which caused the program to essentially play a random move.

SJLIM: fishbait : for crazybird: I think a lot of people condemn IBM for not having Deep Blue play in more matches after beating Kasparov. Is that fair?

CrazyBird: the team was burned out, and the only possible opponent was accusing ibm of cheating...

SJLIM: Tennis : my question for crazybird is what computer program language was deep blue written in?

CrazyBird: as i said earlier, kasparov had his chances, but he blew it.

CrazyBird: tennis was asking me this question.

SJLIM: Yes.. I believe chessbase covered the story of the rematch between yourself and kasparov's agent for those that wish to learn more..

CrazyBird: it is in c, not c++, due to historical reason. the number of lines is in the order of hundreds of thousands.

CrazyBird: the initial dt-2 code is much smaller though.

SJLIM: Yonney : please tell me if DBlue is able to beat kasparov now that he's in his twightlight career?

CrazyBird: i have the number somewhere in the book. don't remember offhand.

SJLIM: Get the book folks! =)

SJLIM: Question: If IBM has no intention of ever letting DB play again, why do you think the evaluation function would still be under NDA, so to speak?

CrazyBird: well, it would be hard, with deep blue distributed between museum(s) and ibm.

CrazyBird: smithsonian is getting one frame. computer history museum might be getting some cards as well.

CrazyBird: it seems deep blue is getting old faster than kasparov:).

SJLIM: unfortunately =)

CrazyBird: next?

SJLIM: there was a question about NDA..

SJLIM: Question: If IBM has no intention of ever letting DB play again, why do you think the evaluation function would still be under NDA, so to speak?

SJLIM: sorry if you had answered it?

CrazyBird: well, i don't have the evaluation function. ibm was keeping the option open, just in case.

SJLIM: I see..

CrazyBird: anyway, i don't have an nda with ibm regarding to the software evaluation function.

CrazyBird: i do have the hardware evaluation function, but that is under nda.

SJLIM: sorry.

CrazyBird: any more questions? or any followup question?

SJLIM: I got hit with a wave of lag.

SJLIM: TheFischerKing : computers always seem to be weak in the endgame phase...why is this? is it a very human phase of the game requiring a method of thinking a machine simply cannot reproduce? do you see this problem being solved in the near future??

CrazyBird: that is my least favorite part of the game.

CrazyBird: there is no way around it. you just do something with the knowledge required.

CrazyBird: lots of special circuits were added in deep blue for the endgame.

SJLIM: Joseph-K : My question for CrazyBird is what have you learnt about your programme given it's play against the world champion?

CrazyBird: i had not figured out how to do coordination squares though.

SJLIM: coordination squares?

CrazyBird: anyway, nasty stuff. part of the reason why shogi is more interesting:).

SJLIM: It will be interesting to see what you come up with in the world of Shogi!

CrazyBird: that is, some king ending, you can draw only if you can coordinate your king with opp's.

SJLIM: I see.. opposition and triangulation! =)

SJLIM: my question is: Murray Campbell uses co-ordinate squares in his Ph.D. thesis extensively -- why were you and he unable to get that happening? was this only due to time constraints?

SJLIM: I assume these are the very same coordinate squares.

CrazyBird: the algorithm for calculating the squares are not easily parallelizable...

SJLIM: julio-cesar : My question for CrazyBird is Did you think that, if the Turing test should be done over a chessboard, in, say 5 to 10 years, you could find computers playing really like humans, in an indistinguishable way?

CrazyBird: anyway, back to your last question. what i learned from the match with kasparov?

SJLIM: opps.

CrazyBird: i need a long rest from computer chess:).

SJLIM: heh

CrazyBird: that is an interesting suggestion about turing test. but it may be hard to do.

CrazyBird: scientists like easily doable experiments. we are lazy, you know.

SJLIM: =)

CrazyBird: my wife is cooking something smelling really good.

SJLIM: I was about to say..

CrazyBird: i may have to leave soon. it is nice to talk to you all.

SJLIM: I think we've answered alot of questions.. but unforunately.. there are so many more.

SJLIM: Thank you so much for agreeing to come back to answer more questions Crazybird. =)

SJLIM: I guess we all look forward to Kasparovs game with Deep Junior.

CrazyBird: you are welcome. yes, that should be doubly interesting now.

SJLIM: I'd like to wish you all the best in your quest to dominate Shogi! =)

CrazyBird: it is just for fun.

SJLIM: Folks. Alot of questions that you may have asked are answered in Hsu's book - Behind Deep Blue: Building the Computer That Defeated the World Chess Champion by Feng-Hsiung Hsu

SJLIM: or in the preview interview.. we will put up a mega transcript of both interviews on ICC as well as on TWIC I hope.

SJLIM: Thank you once again CB Hsu. Enjoy your breakfast.

SJLIM: Thanks to everyone for your participation. =)

CrazyBird: once again, thanks for coming. good bye.

 

logo About Us | Contact Us | Privacy Policy | Terms of Service

© Copyright 1995-2010 Internet Chess Club, Inc. All Rights Reserved.