Greengard's ChessNinja.com

Idiot Savant Computers

| Permalink | 24 comments

We all know that chess programs play brilliantly in tactical complications and not nearly as well in other phases of the game. For openings they rely on databases of human games and professional opening book makers. For endgames they access tablebases, dozens, even hundreds, of gigabytes of databases that allow them to play simple endgames perfectly and improve overall endgame play significantly.

Great strides have been made in recent years in addressing middlegame blind spots as well. Closed positions are still problematic, as are openings with many exchanges. Both of these trends were on display in the second round of the Bilbao Human-Machine tournament in northern Spain. The most notable game by far was Ponomariov-Fritz, which exhibited progress against some stereotypes before reaffirming others.

As this on-site ChessBase report mentions, Fritz had a winning position against Ponomariov before blundering into a lost endgame. (This preceded by Ponomariov forgetting the en passant rule?) A passed pawn was beyond its horizon and it happily won two pieces for a rook before realizing the pawns were too strong to stop. What I don't understand about the ChessBase item is that it misinterprets the start of the game and leaves out some interesting information about Fritz's blunder.

The report says that Fritz sacrificed its b-pawn in a position where it got no compensation and that White's win should have been "just technique" before Ponomariov's blunder on move 31. I'll agree it was a dubious sacrifice, but they don't credit how incredibly well Fritz played the position. Even before Ponomariov's slip, which should have lost, it would have been one of the more remarkable computer games on record. White was left practically without moves and the position was probably not worse for Black before the double blunder. Fritz's play with the h-pawn and on the c-file were very human-like. No less a computer chess aficionado than Garry Kasparov was very impressed with Black's play in this game.

As for the blunder, Fritz has long had trouble with passed pawns, like most programs. One that has far fewer such difficulties is Shredder, the undisputed endgame king of computer chess. I'm always impressed by how well it evaluates endings for a computer, although like all machines it is entirely baffled by blockades. Shredder barely considers the losing move Fritz and Junior evaluate as one of the top few moves even after five minutes. After just a few seconds Shredder drops 39..Bc2?? in favor of 39..Qxg3 or 39..Nc4 and never comes back to the bishop move.

Of course you can always find positions that some programs play well and others don't, but this one is more notable than most. The ChessBase explanation is the calculation horizon, so it's interesting that Shredder does not play this move, even viewing it as a bad, in a few seconds. It doesn't see the full winning continuation, however, for over a minute. The key move 43.a5 is what they all miss at first.


I do wish more of these apps were available for my new Mac. Shredder is, I see - guess it (he?) will be my new sparring partner.

When will Chessbase develop for the Mac? For a Germany company, I'm surprised at how Windows-centric they seem to be.

Mac has far less penetration in Europe than in the US. They have answered the Mac/Linux question many times. It's simply not worth the investment. Lots of work, very few sales. That might change now that Mac is going with Intel, not sure.

I doubt the chipset matters much. It's more about two completely different programming API's. Mac has barely 5% market penetration (according to generous estimates); Linux has even less.

It matters because OS X will become a stronger competitor to Windows and it will have demonstrably better performance. It focuses on Mac as software, and even dual-boot machines should be possible. An increase in market share could make a Mac version of Fritz worthwhile. Also, chess software hits exactly where Intel and PowerPC differ : number crunching. Some chess programs are very much tuned to specific processors.

Hi there,

Just typing this on my wonderful Ibook. My recommendation to Ashish: You can have HIARCS for MacOS under Sigma Chess. Always liked HIARCS the most of all programmes for its style. And there is also Chessmaster 9000 available, with good chess lessons and instructions for beginners and intermediates.
So just leave "the microsoft of chess" alone... ;o)



Market share is the only relevant consideration (for deciding whether to do a Mac port of Chessbase), and it remains to be seen if OS X will make any headway or not. Purely speculative at this point.

If Chessbase were enterprise software, I'd agree on the chipset making a difference, but I don't know if Chessbase is particularly CPU-intensive. It's a database, not a playing engine.

Fruit is also available for OS X, so along with HIARCS and Shredder, Mac users have some very decent choices for strong chess programs. Furthermore, within a year or two, there will probably be emulation software available that will allow you to run Chessbase on a new Intel based Mac with virtually no performance hit.

IIRC, Chessbase did come out with a Mac version sometime in the 90s. It must not have sold very well.

I'm talking about the playing programs. It will be a long time before a Mac port of ChessBase will be viable as long as the port isn't trivial. Fritz is another question, but still far from a priority item. Plus, ChessBase isn't nearly as useful without a playing engine and if it's significantly weaker on the Mac it won't look good.

Market share isn't the only thing or nobody would make software for Macs other than Apple. It's sad in a way because GMs are an ideal Mac audience, just like musicians, artists, and other folks with specialized needs.

Market share isn't everything, but not because of the reason you said. If the total market size is significant, then even a small market share can yield profits, which is why companies do indeed port products to the Mac. But the total market for Chessbase is relatively small.

The picture is complicated because, as you note, for certain products (e.g., graphics software), the Mac market share is much larger than usual.

I can't see Chessbase as being one of those products, though. What does the Mac offer that GMs in particular need? Is the average GM hungry for a Unix kernel?

Simplicity, power, and ease of use, the same thing artists are looking for. But ChessBase software would have to make a huge useability leap to enter that realm. But the number of serious chessplayers in places with significant Mac penetration is too small anyway. If Mac OS becomes a serious alternative to Windows that might change, but even then very slowly.

Please don't tell me I'm wrong and then restate my point, which was that share of market isn't what is relevant, it's number of users. Quite antagonistic!

So hey, I am not a big fan of computer chess, so let me see if I have this straight: computers are better at endings, humans at openings and midgame, and computers are better at open dynamic positions and humans at closed positional situations? I realize this is a generalization, but I was curious. As you can tell, I follow computer chess very little.

It's probably not good to generalize so much, but the basic truth is that winning a chess game is 99% tactics and computers play almost perfectly in these sharp counting positions, while it is exactly in these complex situations where humans have the most trouble. Humans get cut apart by conmputers in the sharp tactics, especially in faster time controls.

I'd say humans are better or equal in most endgames. Only when you get down to seven or eight total pieces on the board do the comp tablebases and sheer calculating power compensate for the lack of logic and the inability to plan. For example, even a weak human can see a 2 vs 1 pawn majority will eventually lead to a passed pawn but a computer has to "see" it all the way to promotion. So they get into trouble with pawns all the time.

Closed positions have the same problems for computers. They can't use the goal-based planning humans use. They have to start at move one and project all the way out. If they can't find a move that is concretely better than others, which is almost always the case in dynamic positions and rarely the case in closed positions, they begin to drift. They rarely blunder of course, but accumulation of small mistakes can lead to a positionally lost position. Happens all the time (though less and less) against GMs. The problem is then turning that into a win, which involves crashing through at some point and letting tactics into the game. Finishing a computer off is far harder than getting a superior position against one, and it takes a very strong human to even get that far these days.

Computers don't play openings, they used human-created and human-tailored databases. If you turn off a computer's opening book it will make very simple, routine developing moves. No real blunders, but no sophistication or understanding of basic pawn principles for the most part. This is because the programmers know they have the opening book to rely on and so don't concentrate on teaching opening principles to their creations. Against a GM the lack of an opening book would be a huge handicap, although again, winning would still be very hard.

I still can't see why GM's need any of the three factors you mention, more than the average person. *Perhaps* there's a case to be made for power (in the brief time window until PC's catch up), but I can't see GM's as needing ease of use or simplicity more than the average user.

No antagonism was intended. I read your "market share isn't everything" comment in the context of the rest of the paragraph, which discussed niche markets for Macs, so I assumed that was what you were talking about.

In my defense, I don't see any mention of the "number of users" issue, even if that was what you meant. In any case, the last issue of size-of-pie-vs.-slice-of-pie isn't relevant unless overall sales of Chessbase increase significantly.

Computers and endings are tricky. There was a point where the computer was absolutely awful in endings but the the tablebase came along which made it play absolutely perfectly in endings with less than 6 men. However, it can still get into trouble with passed pawns as this game Ponomariov game shows or with many fortresses. Can anyone think of anything else in which the computers are noticeably weaker? Anyway, any position which is very tactical is where the computer shines. So obviously the computer will be good in tactical endgames(studies being an extrem example) but maybe not so strong in endgames that are not very tactical but involve long term planning or strategical thinking. Just some quick thoughts.

I believe the time has come for competitive chess software to be allowed only a self-created (games and stored evaluations the program has actually "played") opening book...for that matter no endgame tablebases....the only exceptions would be in slow time control games where human opponents would have access to openings databases and the endgame tablebases, such as the Advanced Chess played in Leon.....

Openings books were necessary in the infancy and youth of chess software; now it is ridiculous for mere humans to play an opponent (under modest human time controls) with virtually perfect, virtually instantaneous recall.

As for the advantage (rules violation?!) of the software "analysing on another chessboard", that is another subject.

A beneficial potential effect of tossing out the openings books is that good programs may produce meaningful advances in opening theory...maybe 1 g4 is better that humans have thought!

I don't think that it's true engines have problems with simple 2 vs 1 pawn endgames. The evaluation function can be implemented to include scoring for basic "winning structures" such as a decisive pawn majority on one side.

One aspect of the game that engines still continue to disappoint in is evaluation of demolition sacrifices. You sacrifice on h7 or g6 or something and the engine runs long lines where it just shuffles its pieces while the attack proceeds to a winning conclusion. The engine thinks it's up a piece for a pawn and fails to see that there's no defense 20 ply down the line. This is how the Nemeth demolition strategy succeeds.

Another thing that engines miss is permanently trapped pieces. Run through some games in the Frankestein-Dracula Vienna and see how the evaluation is at first complete nonsense because "White is up a rook".

"GMs are an ideal Mac audience, just like musicians, artists..."

Hi Mig,

Though his words are often dismissed, I always liked Kramnik's comparison of chess with painting. You too?
Remember what Kasparov already said about chess programs in the eighties: "They lack intuition!". In painting you need good skills, but it is nothing worth since you do not feel from the very inside what is the right thing to do with your work.



I was more referring to GMs as "special needs" users in the same way as artists. They need a specialized suite of coordinated tools more than broad compatibility with the rest of the world with, say, office documents.

Chess has long been called an art, at the very least as part art, science, and sport in different measures depending on the player in question. (Stereotype alert: Lasker=sport, Botvinnik=science, Tal=art)

It would have been nice had Kramnik been talking about the same sort of chess/art relationship as Tal. But where Tal was referring to aesthetic beauty and inspiration over the board, Kramnik was referring to his right to not create anything at all if he didn't feel like it.

Unfortunately I think Kramnik has come to believe this himself in the past few years and this self-inflicted alienation from the game and its fans has harmed his play. You can't play something as intense and involved as chess without really caring about the result and not have it affect your performance.

What an excellent game Rustam had against Fritz, in all Shipov awarded 13 exclaimation marks to him. He certainly is the most exciting Elite player alongside Topolov.

Kasimdzhanov seems to be playing the best against the computers, even though Ponomariov is the only one with a victory.

What is the performance rating of the idiot savant computers?

The average rating of the humans is 2675 ... An 8-4 win ... some back-of-the-envelope scribbles ... I'd say the progs performed at about 2810.

Performance rating is 2793 I think. Which is good of course but doesn't make them better than the top humans. So the report from chessbase where they suggest computers should start giving pawn odds is pretty insulting. Also I think if a top gm would specialise 100% in playing computers he would beat them with a big + score as computers still have many obvious flaws.

What is the performance rating of each idiot savant computer individually?

Twitter Updates

    Follow me on Twitter



    About this Entry

    This page contains a single entry by Mig published on November 22, 2005 10:01 AM.

    More FIDE-Kramnik was the previous entry in this blog.

    State Champs Finals is the next entry in this blog.

    Find recent content on the main index or look in the archives to find all content.