The 13th World Computer Championship was held last month in Iceland, but you might easily have missed it. Despite a great deal of organizational effort and a fine official site, news coverage was sparse. And there was news in Reykjavik, thanks to two relatively unknown amateur programs finishing in the top two spots ahead of the usual suspects.
The American program Zappa scored an impressive 10.5/11 in the round-robin for clear first. The open source program Fruit from France was second, two points behind. Shredder, which collects these titles, was tied for third and defending champion Junior was equal 5th-6th. As with human chess, a single tournament win doesn't say anything definitive, but it's an intriguing result nonetheless. I mentioned the event and how you can use some of these programs in my latest column at ChessCafe.com. (PDF) Shredder programmer Stefan Meyer-Kahlen has a diary of the event with photos.
ChessBase, publisher of the top commercial programs Shredder, Junior, and Fritz, covered this event exactly not at all, quite a switch from previous years when the WCCC often received daily coverage. Basically they feel the publicity and scientific returns from comp-comp events have diminished to the point of negativity. The small but vociferous computer chess community spends a lot of time bashing ChessBase, much the way the computer geeks bash Microsoft. (Note that they had decided not to cover the event before it started (and didn't enter Fritz), so the lack of coverage is not about the relatively poor results of Shredder and Junior.)
I did a quickie email interview with Anthony Cozzie, the programmer of the new champion. Most of the questions are evident from the answers, so I've skipped them where possible. Here's a pic he sent of himself with the trophy.
I am a graduate student in computer science at the University of Illinois at Urbana-Champaign (I live in Urbana). This is my first year. I already have an MS in Computer Engineering from Carnegie Mellon in Pittsburgh, home to Hitech & Deep Thought. I managed to take a class in parallel computing from Nowatzyk. I found him a very interesting guy, except that his voice put me to sleep. In my spare time I play basketball and watch anime (I speak some Japanese as well).
I began the program during the spring of my Junior year at Carnegie Mellon (2000 or so) and have worked on it ever since. As a computer scientist I find the parallel search the most interesting part, but I have also worked on the chess knowledge factor. About 2 years ago, just after CCT6.
[How much has it improved since the last version, which you said was significantly weaker than the top commercial programs?] I would say that the program is quite a bit stronger, but to be honest I have no hard proof. I have never run a match of (say) 100 games vs Shredder, nor have I ever tested my parallel speedup on more than 3 positions. I tend to 'fly by the seat of my pants' so to speak :) This has gotten me in trouble from time to time, though.
[How much do you pay attention to other programs?] Quite a bit. Its simply much easier to start a Zappa-Shredder match than a Zappa-Kasparov match to see whether the program plays better chess or not. I think that Zappa and the Erdogan [opening] book are simply a beautiful fit. Erdo's book simply needs a deep searching aggressive engine, and Zappa fits the bill. I myself am probably 2000 USCF or so, so yes, I bug any strong chessplayer I can find :) It is also something of a problem living in the US where there aren't as many strong chess players.
I leave the book work to Erdo. You'd have to ask him how many hours he spent on it, but I suspect it's in the thousands. I think the book is extremely important: just look at the emphasis top GMs place now on the opening. Good programs are now of similar strength, and the opening is just as important for them. This is especially true because to win open tournaments you have to get an insane score, and that means winning a lot with Black. There simply aren't many pushovers in computer-chess these days, so you need a good Black opening book.
what is the hardware Zappa runs on? A multi processor cluster? Windows or Unix? Any custom hardware? Programmable logic (FPGA/CPLD)?
Erdogan [opening] book are simply a beautiful fit. Erdo's book --- Will someone please help me here? What is he talking about when he refers to Erdo's book and Erdogan opening book? A cursory internet search for it didn't turn it up. Pardon my ignorance.
Go Illini!
Erdogan is just the name of the man who created the opening book used by Zappa. His name is Erdogan Gunes (must be Turkish) and he also worked on the book for Hydra. He's been in computer chess for a while.
Why Zappa won the tournament ?
Did he play better chess [that is to understand the positions] or simply could calculate more variations [or simply because of better opening "preparation"]
It could be interesting if program's creator could answer this question..
Fruit is heavily tested and is about the same strength as Shredder 9. It's freeware. Equally strong is Turk II, which is also free. Spike beat Shredder 9 at Chess690 recently, and it too is free to download.
I don't know how its possible, but the good days are over for the Chessbase team of programmers. Six months ago Shredder 9 was the gold standard for all computer chess programmes. Its position was dominant. Neither Fritz nor Junior could keep pace. Now these amature programes come from nowhere and outperform the Chessbase stable. It goes without saying that all perform beyond human abilities, so there's no question that chess engines have lost commercial value. They will become part of other chess products, such as training cd's etc.
Perhaps the challeng now is to write chess engines that improve human play. What would that be like?
In my previous post I said "Turk II" in error. I should have said Toga II.
Although, of course, "Turk" would be a name with acceptable pedigree.
But you're only talking about the engine. There are probably 20 free engines that would crush any human rated below 2500 equally well. The strength of the engine is something you tout when you have it, but there has been little difference at the top for a long, long time from a human perspective. Yes, a few programs stayed ahead of the pack by a small margin for a few years, but engine strength is like deciding which car to buy based on maximum attainable speed. A cool factor for real geeks, but useless to 99.9% of the consumer population.
ChessBase and its Fritz-family products are relevant primarily because of the interface. This has been true for years. Cool graphics, user-friendly tools, training and handicap features, print and online publishing, database access, automated analysis, etc. You can't download those things for free. Having Fruit or Zappa running in a free interface is definitely cool, but Crafty has been strong enough to substitute as a powerful freeware engine for years now. Shredder and Junior finishing behind some amateur programs in an 11-round tournament is noteworthy, but hardly a sea change.
Unless you're really into comp-comp chess, having an engine rated 2800 is no different from one rated 2700 or 2600. They all play humans about the same because they play so differently from humans. ChessBase has already said they are emphasizing interface and human training in the upcoming Fritz 9. Comp-comp play is still somewhat interesting, but it's been a largely academic exercise for quite a while now. That's not an insult, just the nature of a mature, increasingly crowded, and now largely open-source area. They make small, incremental advances that have almost no practical market utility now that they are all so incredibly strong.
Making them smarter so they can understand and coach typical human mistakes is the new challenge, and in many ways a tougher one.
Regarding GUI, freeware Arena is a very nice GUI for any UCI/winboard engine, including Shredder UCI. No Fritz or Junior support, of course ;-)
And I don't think it's being updated anymore, but there's a freeware database called SCID that was going well for a while.
http://scid.sourceforge.net/
Mig,
I know you're a strong player (you don't say so yourself, but by ordinary standards you are). You may not find any value in a product such as Bangiev "Square's Strategy" CD's. Personally, as a weak player (by any normal standard), I find them good, although difficult to use because of their very ... em ... idiosyncratic use of symbolism.
It seems to me that Bangiev has a vision of chess instruction that could be rendered by chess programmers. Bangiev's method is entirely concerned with the geometry of the board, and therefore does not involve what might be called higher concepts (such as those lovely neologisms from Hans Kmoch or aphoristic phrases of Nimzowitsch).
Do you think "Frangiev 1.0" would be a viable product?
I've found quite a few of the ChessBase training CDs relevant, although I haven't really worked with any of them extensively.
No doubt you could put a marketing spin on it, and engines and programs in the past have been touted as playing like so-and-so or adjustable to specific human styles. But computers don't use geometry in generating moves. It's just that their moves often reflect the natural geometry of chess more than human chess because they don't have any problems seeing things like long-distance triangulations and similar stuff difficult for humans to visualize.
It's humans who need crutches like geometry or square theory to aid our dependence on patterns and inability to calculate without prejudice. After years of doing the newsletters and working with (and playing against) 1200-1800-rated players, I have a long collection of positions and tactics that are quite simple (i.e. few pieces, things that don't require a lot of calculation) but that are very difficult to find because they aren't typical themes.
Anything that kicks you out of your mental routine a little can be helpful. Keeping your thinking fresh stops you from getting bored and predictable, which is the death of improvement unless you have an incredible work ethic or a masochistic streak.
Dear Mig,
there is some inconsistency in your remarks. The point is not what chess engines presently do very well, but what they will do in future to remain marketable. I do not question your understanding of how chess engines generate inevitably strong moves. And your point that these moves reflect human perceptions (geometric or otherwise) of good chess play is well made. But the challenge to chess programmers, to "understand and coach typical human mistakes" (your words) is, indeed, new and of a different order. It is the challenge to reflect or comply with, doubtlessly, a range of human though processes. Or, one might say, to become 'intelligent' and not just 'excellent' in their play. One cannot ask the impossible. Human play is inimitable (for a computer), because it is emotional and sometimes personal. What can one ask, then, of chess programmes in the future? Surely, the future market depends on answering this question, rather than on repeating the successes of the past. Therefore, it is not valid to answer that chess programmes do not need 'crutches' to generate near perfect chess. Their legs must be broken! Make them dependent on crutches, so that we may move haltingly and slowly forward together.
I was answering two different questions. One was about computers using Bangievian geometry. You seemed to imply that this was a natural fit for computers and I was only pointing out that they don't think this way. Not an endorsement, just a statement of fact.
In the short run (the next few years), dramatically swapping speed for knowledge will make a program a little weaker (relevantly against other comps and super-GMs) but more useful to the market if done well. That is the sacrifice. If people are obsessed about comp-comp scores and which program is thusly proved stronger, it makes progress in a more interesting area tougher.
The best example of this is how bad most handicap settings still are. I've never seen even a remotely good simulation of a 1400 or 1800 player. Comps just play such a different game that the mistakes they make are entirely atypical of human chess for the most part (just like many of their good moves for that matter). This also means it can't be set to allow the human opponent certain opportunities natural for that level. If a more intelligent program can be made into a useful handicapped training partner for players from beginner to master, that will be worth far more than another 50 comp-comp Elo points.
Agreed!
The largest market, then, is for chess programmes which mimic accurately the play of, say, the strong club player. (Provided the club is in the US or UK, rather than the Ukraine -- see http://www.chessbase.com/newsdetail.asp?newsid=2615)
I wonder if progress will be made in the Computer-Herman... whups, sorry... Computer-Human interface by considering more than alpha-beta and evaluation functions. I haven't looked into computer chess for a long time so I should say "will be made or has been made".
My particular thought is to pull more information out of search trees, for example: Suppose search trees were represented topographically or in an equivalent mathematical sense where a board position's 'elevation' corresponded to evaluation score: As evaluated in place, not propagated back from deeper in the tree, or at best propagated back from a short search horizon. (This tree exists in parallel with the usual tree... except it uses a lot more memory I suppose because it has to be examined as a landscape.)
Suppose that a forced line exists that does not evaluate badly for many moves, a peninsula of topography -- reasonable looking positions -- that eventually leads to cliffs on all sides. Suppose further that a computer can choose a move that leads to a position from which there are many such peninsulas of death (for the Herman) and only one correct method of avoidance. Might it not use this situation to supersede a marginally better move, essentially betting on the human's odds of finding the way out?
(Incidentally I imagine one might use the usual full-up search tree to prune this topographic tree to smaller size.)
Usually strategies of this type are overcome by branching factor; the game is just too complex to rely on finding peninsulas of death. Hence you've paid a huge computational cost to look for something you'll seldom or never find.
But take this a step further and generate many topographic trees, each with a different search horizon. This simplest just uses a static eval function to assign scores to each board position. Next tree uses a depth-1 search from each position and then depth-2 and so forth. Stacking these topographies you'd see some board positions (let's say nearer the base) that change their evaluation as depth increases while others remain fairly stable and banal with depth. I wonder if one could come up with a little vector space that characterizes this variability-with-depth and steers the game towards positions accordingly, to match what a person wants in training for example.
The idea is not to mimic our play but to pull more information out of search trees in attempting to make computers more relevant. To Hermans.
Hi Everyone
I need a GUI which works on OS X for my UCI engine. Any recommendations here?
Sigma Chess (www.sigmachess.com) is the only UCI interface I'm aware of for OS X
The Zappa Homepage:
https://netfiles.uiuc.edu/acozzie2/www/zappa/
The Diep3D Website (where the commercial Zappa WCh. version will be published):
http://www.diep3d.com/
For people who read german: CSS Online has covered Zappa versus GM Ehlvest, Zappa's Ranking in the CSS rating list and the Computer World Championship in Reykjavik in the News:
http://www.computerschach.de/index.php?option=com_content&task=category§ionid=1&id=1&Itemid=109
MIG: "There are probably 20 free engines that would crush any human rated below 2500 equally well. The strength of the engine is something you tout when you have it, but there has been little difference at the top for a long, long time from a human perspective. Yes, a few programs stayed ahead of the pack by a small margin for a few years, but engine strength is like deciding which car to buy based on maximum attainable speed. A cool factor for real geeks, but useless to 99.9% of the consumer population."
I think it is wrong to assume that what practically every consumer is looking for is an engine that is fun or instructive to play against.
If people want to match an opponent that plays like a human on about their own level, it is easy to simply log on to a chess server and play against, errr, actual humans. So you might say that attempts to make an engine mimic a human is useless to 99.9% of the consumer population.
The truth is probably somewhere in between. I would just like to point out that many users like to have strong engines because they are good for saving time when wanting to "find the truth" about a position. They might not be interested in playing against an engine at all, and they may also not be engine geeks who cheers for engines going 20 elo points higher on engine charts. They just want a close to state of the engine so they can find the "truth" about a position as fast and well as everybody else.
Why would you NOT want the strongest and fastest engine to help you in that quest?
I think the line of reasoning that people aren't/shouldn't be interested in stronger engines because they are already crushed by practically every engine already is somewhat similar to saying that people shouldn't care for 6 piece tablebases because they are already sorely beaten in 6 piece endgames by engines merely using 5 piece tablebases. I think both lines of reasoning are flawed to at least a good portion of the consumer population, not only the engine geeks.
I agree.
There is no such thing as a too strong engine when you are analysing your tournament games.
Yes, but currently the difference between the top engines is entirely null and void when it comes to analyzing human chess. They each see some positions a little better than others, and occasionally come up with different moves in the same positions, but there is little if any qualitative difference overall.
An engine that is a lot smarter would be MUCH more useful in analyzing human games even if scored 40 Elo points less in comp-comp chess than the top comp-comp programs. That's my point. You are using the word "strong" as if there is a universal measure. My point is that comp-comp results do little to distinguish human utility among the top 10 programs.
What do people think of programs playing in national champs and other events? Seems to me a total disgrace? Chess tourns should be a human contest. But by all means have special promos events with gimmicks like having a compter entry - but not in formal events.
There is only one real reason why chess programs do NOT participate in national champioships and other "formal" events:
Because they are strong, annoying, nasty rivals! :-)
this wasn't newsworthy on chessbase.com hmmmmmmm?
Mig is right that the value is in the interface and user-friendliness of Chessbase products. But this crushing loss at the hands of an unknown program has to be a marketing problem for them. Just look at how they position Shredder on chessbase.com's on-line shop:
"For computer chess experts Shredder is the number one choice. Nobody can ignore its amazing seven computer chess world championship titles, won at Jakarta 1996, Paderborn 1999, London 2000, Maastricht 2001, 2002, Graz 2003 and Tel Aviv 2004."
Not much emphasis there on the interface! Just pure, raw POWER.... Rightly or wrongly, people buy Fritz and Shredder because they think they are the best. And that - in their minds - justifies the 50 USD (or whatever) investment. If people realize that programs available for free are more powerful, then it's clearly going to put a dent in sales of Chessbase engines.
Yes, this focus on power makes no sense for 99.99% of the people who buy a chess program. But how many people who buy a Ferrari really need to go from 0-100 in 8.8 seconds?
I suppose how serious this is really depends on how dependent Chessbase is on their engines. I guess that is why they have moved into on-line chess, with a subscription model. It would be interesting to see how they earn money - break out revenues by product. But private companies have no need to publish that info, unfortunately.
I wonder if Chessbase is considering stopping all investment in Fritz, Shredder, etc., and just selling the interface? They could bundle it with one free engine (Zappa? Fritz Open Source?), but allow multiple engines to be used with the interface. But again, even though the value truly is in the interface, I'm not sure how many people would be willing to pay for "just" the interface....
--JC
I could be confused on the issue, but I'm pretty sure as it stands chessbase does not own shredder, junior, and fritz yet it packages all three for sale so I believe they've significantly pre-empted your suggestion JoeChristmas. Also, remember that the computer vs. computer results do not necessarily indicate engine strength. The fritz group has specifically mentioned that they are reducing the amount of research they spend adding anti-computer strategies into fritz and adding more human useable features.
I don't have any particular problem with Chessbase, but I do wish they would cover the WCCC, for the same reason they cover human-human tournaments: drama and interesting high-level chess. Of course, since Chessbase products stopped winning they have stopped covering ;) I don't blame for a good business decision, but that doesn't mean I have to like it.
I think the primary benefit of a stronger engine is analysis. One of the viewpoints that a lot of strong chessplayers have is that 'oh, computers are only good at finding tactics'. And sure, any chess program will do well there. But of late computers actually play pretty well positionally. I mean, look at Zappa-Junior. You could easily think a strong GM had White there. Zappa played solid positional moves the whole way. So I think that the latest programs can make positional suggestions as well as tactical ones.
Unfortunately, I do agree that they are a long way from explaining WHY a move is bad or 'I see this in a lot of your games, you should fix it' or such things. I can say that the commercial version of Zappa will include some weakened engines so that we humans can have fun playing and not lose *every* game.
cheers,
anthony
C'mon, everybody knows that Zappa has No Commercial Potential.
Frank Zappa Rulles!!!
Where Can I get zappa 2 chess engine? I want to test it against Fruit 2.2.1 and Toga IIa.