Author
|
Topic: Man vs Machine: Kasparov - Deep Junior (Read 828 times) |
|
rhinoceros
Archon
Gender: 
Posts: 1318
Reputation: 8.06
Rate rhinoceros
My point is ...
  
|
 |
Man vs Machine: Kasparov - Deep Junior
« on: 2003-01-28 12:42:38 » |
 |
In the first game of a new "Man vs Machine" chess match, Kasparov beat Deep Junior in 27 moves. Here is the game:
http://www.chessbase.com/games/2003/x3d1.htm
Comments from chessbase.com:
http://www.chessbase.com/columns/column.asp?pid=159
<snip>
Kasparov crushed Deep Junior in just 27 moves. A powerful opening novelty led to a dominating position. Junior sacrificed the exchange to try to escape the pressure, but Kasparov stayed in control and easily converted his advantage.
<snip>
Kasparov arrived to the board and played 1.d4. In recent years Kasparov usually only plays this instead of his customary 1.e4 when he has something special cooked up in the opening kitchen. As things turned out, that was again the case.
A Grandmaster might have been alerted by Kasparov's switch and tried something off the beaten path, but Deep Junior couldn't do anything other than follow its opening book right off a theoretical cliff. A human's alarm bells really would have been ringing when Kasparov played 7.g4, entering the sharpest lines of the Semi-Slav Anti-Meran. This lunge is credited to US Champion Alexander Shabalov, although it was first played by his Latvian countryman Alexei Shirov back in 1992.
<snip>
Kasparov clearly wasn't going for the cautious style of play generally recommended against computers. Kramnik beat Fritz in Bahrain in games two and three by trading the queens and exploiting positional advantages and exceptional endgame technique. In other words, he won like Kramnik wins against humans. Kasparov won with an opening improvement in a sharp line and then blew Junior off the board. He won like Kasparov wins against humans. Opening preparation, tactical control, resignation, reboot.
This is an interesting twist after a decade of anti-computer chess debates. Kramnik pulverized Junior in Dortmund 2000 with typical stonewall chess that computers didn't understand. In Bahrain last year he told me that the programs had already improved beyond that point, and that to beat them there were no more tricks, just good chess. So now we have computers that play like humans, and with that we have computers that lose like humans.
After the game an ebullient Kasparov told the ecstatic audience about how he and his trainer had discovered the major improvement in this line (13.d5!). There were various opportunities for Black to diverge before this into sidelines, although it is almost certain that Kasparov had something nasty prepared in all of them.
<snip>
Junior took so long on its ninth move that there was a report of a technical problem by the onsite commentators! (And I relayed this online.) Junior's co-programmer Shay Bushinsky arrived on the scene to put these fears to rest. (Programmer Amir Ban was operating Junior in game one.) Bushinsky explained that it was not unusual at all for Junior to use a lot of time when it encountered a negative shift in its evaluation. Since it had just come out of its opening book to the unpleasant surprise of finding itself much worse, it thought at 20-minute think was warranted!
The game was still in Kasparov's preparation, but Junior was out of book and not happy at all. It didn't get any better for the Israeli program. Kasparov planted a knight on d6 and had a completely dominating position. After the game he said, "All my life I've believed that a knight on d6 was worth a rook, and this game proved it!" Junior must agree with that maxim because a few moves later it sacrificed a rook for the knight!
<snip>
When a computer gives up material you know it's either in serious trouble or about to announce mate in 12 on your sorry human butt. Garry spent 10 minutes making sure it wasn't the latter before grabbing the rook. His trusty trainer, GM Yuri Dokhoian, said during the game that this type of move was not unusual at all from Junior and that they had seen it make many similar sacrifices in their preparation. After that Kasparov spent some time consolidating his position while Junior futilely looked for counterplay. With 24.b3! Kasparov forced the trade of the heavy pieces and it was all over. On move 27 Amir Ban resigned for Junior.
Many thought Garry might sac his queen with 25.Qxc8+ Bxc8 26.Rxc8+ Nf8 27.Nxb3 to finish with a flourish. Those flourishes are exactly the things that get human killed in winning positions against computers. It does look like an easy win as well, but leave a computer's queen on an open board and you are just asking for it to find a bizarre perpetual check or insane tactical trick that forces mate. And you can't argue with a safer move that causes resignation two moves later.
<snip>
There was a huge cheer when Ashley announced that Deep Junior had resigned. Before the game around 80% of the audience had voted that Kasparov would win the match and it looks like at least that many also want him to win. In my ChessNinja.com poll, 64% predicted a Kasparov match victory. After the game Kasparov reminded everyone that he had also won the first game of the Deep Blue rematch before going on to lose 3.5-2.5. And that Deep Blue had won the first game of their first match and had also lost the match.
<snip>
|
|
|
|
|
rhinoceros
Archon
Gender:
Posts: 1318
Reputation: 8.06
Rate rhinoceros
My point is ...
|
|
Re:Man vs Machine: Kasparov - Deep Junior
« Reply #1 on: 2003-01-29 04:23:41 » |
|
Man vs Machine, Game 2
A draw. The score is now Kasparov 1.5 - Deep Junior 0.5.
Kasparov, playing black, gave up a rook for a bishop and had good winning chances. But an inaccurate move in his 25th move (Queen to a1 check, instead of Pawn to f4) allowed Deep Junior to find a way out by giving up his Queen.
Here is the game.
http://www.pocketfritz.de/x3dflash/english
More commentary later.
|
|
|
|
|
rhinoceros
Archon
Gender:
Posts: 1318
Reputation: 8.06
Rate rhinoceros
My point is ...
|
|
Re:Man vs Machine: Kasparov - Deep Junior
« Reply #2 on: 2003-01-30 05:43:24 » |
|
[rhinoceros]
A very interesting evaluation of the Kasparov - Deep Junior "Man vs Machine" match so far:
http://www.chessbase.com/columns/column.asp?pid=160
Real Chess against a Virtual Opponent
After an exciting draw in game two, Garry Kasparov maintains the lead in his six-game match against Deep Junior 1.5-0.5. The first two games of the Kasparov-Deep Junior match have shown us some excellent fighting chess.
There are several noteworthy things apart from the high quality of the games.
1) Kasparov is not avoiding tactical complications. It has become almost conventional wisdom that allowing complications with queens on the board is a sure death against computers. In game two, Kasparov played a Sicilian Defense, sacrificed the exchange, and came very close to a winning position.
2) Deep Junior has so far passed the chess Turing test. Its play has been almost completely indistinguishable from that of a human master. I'm not sure it's really playing like a human Grandmaster so far since it has reached losing or nearly losing positions in both games by move 25. But it hasn't made any obvious computer-like moves.
3) In both games Junior made a mistake on its first move out of its opening book. 9...e5 in game one and 12.Nd5 in game two were weak moves that ignored important strategic factors in the respective positions. In general, if a computer is making mistakes right out of its book it shouldn't be getting into those positions to begin with because it doesn't understand them.
The first two games have confirmed much of what we saw in the Kramnik-Fritz match in Bahrain. The top humans are significantly stronger than the top programs at classical time controls. Before we start talking about preparation and anti-computer play and opening books, we should remember that Kasparov is the strongest player in the history of the game. It shouldn't shock us too much when programs that couldn't beat any Grandmaster 10 years ago can't beat a 2847 player today.
Junior scored 50% in the Dortmund (human) supertournament in 2000, a phenomenal result at the time. Kasparov would have been expected to score at least +4 in that same tournament! It's only two and a half years later and it is unlikely that Junior has improved over 150 Elo points to Kasparov's level in that time. It's hard to track increases in program strength over time against humans because they don't play enough games, but after looking at various events over the past five years anything more than 30 Elo points per year would surprise me.
Discussing this with ChessBase director and computer chess guru Frederic Friedel, we surmise that today's top programs play consistently at a 2500-2600 level of chess quality. The difference is that they instantly and mercilessly punish every human mistake and almost never let a winning position slip. This near-elimination of the margin for error pushes their practical performance up toward the 2800 level.
Of course in many tactical positions computers will play better than any human can ever dream, but that doesn't mean a strong human can't keep up even during heavy tactics. Grandmaster chess rarely comes down to pure tactics for more than a few moves, unless it's a Shirov-Polgar game. (Not for nothing has Judit Polgar been called "our ideal opponent" by ChessBase.)
The problem for the humans is maintaining that high level of play throughout a match. Kramnik cruised to a 2.5-0.5 lead against Fritz while playing a fantastic level of chess. He collapsed in the second half of to finish in a 4-4 tie. One game was lost on an elementary blunder in a difficult position (which Fritz played excellently), the other on an uncharacteristically emotional piece sacrifice that Fritz immediately saw for a blunder.
I'm not trying to split hairs. A performance rating is a performance rating and I don't want to make excuses for humans who blunder, because we all do. Fritz drew an eight-game match with Kramnik, period. If Junior wins a game because of a Kasparov blunder, that's fair. However, it is relevant to distinguish between the overall quality of the chess and the results for those interested in the debate about computers killing chess or becoming invincible in the next few years. When the top progams can play moves like Fritz's 12...Bf8?? in game two against Kramnik or Junior's 17...Rae8?? in game one against Kasparov, we still have a way to go before we have to hang up our pawns and take up something computers aren't good at yet, like ice skating.
Kasparov's match is two games shorter than the Kramnik-Fritz match. But one blunder could set off a psychological chain reaction that could be disastrous. Right now Kasparov is riding high after outplaying Junior in two fine games.
In the second game Kasparov sacrificed the exchange for a strong attack. Junior went pawn-hunting with 21.Ra3 and was close to a second consecutive loss after Kasparov gave up a rook in order to mobilize his queen and central pawns. Kasparov was very upset at the board soon after playing (diagram) 25...Qa1+. He realized a few moves later that White could force him to take a perpetual check draw with a surprising queen sacrifice. In the press conference he said that he thought he had a forced win with the queen check, but that in hindsight 25...f4 would have been stronger.
Junior co-programmer Amir Ban quotes someone from Kasparov's team as saying afterwards that 25...f4 26.h3! was also a draw. We can't be sure yet what Junior would have played. My DJ 7 wants to play 26.Nf1, which would allow black a very strong attack after 26...e3.
Things were excellent at the New York Athletic Club site away from the board. A new media area was set up and the commentary room was switched around to allow for a giant display of the board and a camera directly on Garry shown in X3D! Every trademark Garry grimace, frown, and head shake was larger than life. If you think he's intimidating in person or in photos, you should check him out when he's four meters high and popping out in three dimensions. Forget battling against computers. Coming soon to a theater near you, Kasparov versus Godzilla in 3D!
|
|
|
|
|
rhinoceros
Archon
Gender:
Posts: 1318
Reputation: 8.06
Rate rhinoceros
My point is ...
|
|
Re:Man vs Machine: Kasparov - Deep Junior
« Reply #3 on: 2003-01-31 10:19:20 » |
|
[rhinoceros]
Kasparov lost the 3rd game of the match against Deep Junior. The score is now a tie: 1.5 - 1.5.
Kasparov had been dominating the board for most of the game with a spectacular all-out attack, but failed to keep control after a careful defence by the computer and found himself in a lost endgame.
In this match, Kasparov seems to play in a way contrary to what is known about how one should play against a computer: A style demanding a lot of calculations. This somehow reminds of one of his idols, Alexander Alekhine, one of the greatest players of all time and one of the most controversial figures in chess history, who often tried to outplay his opponents in their own style: Playing strategically against a strategist and tactically against a tactician.
Here is the 3rd game:
http://www.x3dworld.com/Entertainment/chessMVM/DGT/Game3/Game3.html
And here is a review of the game:
Deep Junior strikes back
http://www.chessbase.com/newsdetail.asp?newsid=766
Actually it was Garry Kasparov, who was dominating during most of the game. Then he let his advange slip, and just when he had resigned himself to accepting a draw he overlooked a sharp continuation which handed the game to his opponent. With this surprise victory Deep Junior has equalised and filled the entire match with new tension. How did it all happen? Here's a full illustrated report on game three.
Deep Junior has tied the match 1.5-1.5 by winning game three. In another exciting game Kasparov took the battle to Junior's king but couldn't land a knockout punch. Junior bobbed and weaved like Muhammad Ali and slowly equalized the game. Finally, just when it looked like Kasparov would try to force a draw, the human blundered and lost almost instantly. Kasparov had missed a spectacular checkmate variation that caused him to lose another pawn and the game.
This is exactly how Grandmasters most often lose to computer programs. They get excellent positions and then watch them unravel against near-perfect computer defense. Junior was in trouble for the third straight game out of the opening. Kasparov, playing white, tricked Junior back into the g4 lines of the Semi-Slav (from game one) through a sneaky move order that got the Israeli program out of its opening book.
Then Junior began to play the sort of odd moves that have given its computer opponents so much trouble in the past few years. Junior won the world championship last year with the same risky, positionally suspect chess. It seems to know just how close to the edge it can go without falling off. In many of its games against other programs you see its opponents get very positive evaluations only to see them fall slowly as Junior's evaluation proves more accurate.
No human or computer would argue with the fact that Kasparov's position was superior for most of game three. As always, the problem was turning a "winning" position against a computer into a win. Those quotes illustrate the problem. In a perfect world, with perfect play, it is likely that Kasparov's position after 13.e4! was winning. In the world we live in, with a supercomputer like Junior as black, it was not.
After the tactical complications slowed down on move 21 Kasparov was down a pawn but had a lead in development and considerable pressure against Junior's king. "Fritz 8" was giving a half-pawn plus to White despite the pawn minus. But these positions are almost impossible to win against a strong program. Pawn play was not a factor. Long-range planning was not a factor. It was wide-open piece play with king safety issues for both sides. This sort of thing fits a computer like a tailored Armani suit. A human will never win these positions against a computer. NEVER. You can only lose. It simply sees everything. Unless you already have a forced win on the board it is time to start looking for a way to draw before you fall into something nasty.
This is what happened to the world number one. Kasparov's initiative slowly faded away against Junior's precise defense. Then he realized it was time to look for that draw, only he looked in the wrong place. 32.Ng6+ does not appear to be the forced draw the commentators believed it was, but it was definitely better than Kasparov's 32.Rh5??, which lost another pawn and the game to a brilliant tactical shot.
At the time we thought Kasparov was trying to play for a win. After the game he said he thought the rook move was the easiest way to force a draw. After the apparently forced 32...Qxd4 33.Rxh7+ Kxh7 34.Qxf5+ is a perpetual check draw. Unfortunately for humanity, in the diagram 32...Nxd4!! is a winner.
This seems impossible because of 33.Ng6+ Kg8 34.Ne7+ and it looks like black has to take a repetition. But 34...Kf8! and now if 35.Rxh7 Nb3+!! ouch, it's mate! This is what Kasparov missed with 10 minutes on his clock. 36.Kc2 (36.axb3 Qd1#) 36...Na1+ 37.Kc3 Qd2+ 38.Kc4 b5+ 39.Kc5 Qd6#)
Kasparov tried to bail out with 35.Nd5 but resigned after 35...Qg7 36.Qxd4 Rxd5 0-1. A sad but all too common fate. A marvelous performance ruined by a moment's inattention against a beast that never sleeps. (Curiously enough I wrote about just this sort of thing in Mig on Chess #185 the day before this game was played.) Credit is due Junior for surviving what looked like a hopeless position after just a dozen moves. And congratulations to Amir Ban and Shay Bushinsky for their first win against the world's number one player!
Kasparov was indignant in the postgame press conference. He felt that he had completely dominated the computer in all three games, yet only had 1.5 points and a drawn match to show for his efforts. Now he has two rest days to recover from this devastating loss. Junior has white in games four and six, but I still think the match is a toss-up considering how well Kasparov has played overall so far.
|
|
|
|
|
rhinoceros
Archon
Gender:
Posts: 1318
Reputation: 8.06
Rate rhinoceros
My point is ...
|
|
Re:Man vs Machine: Kasparov - Deep Junior
« Reply #4 on: 2003-02-03 09:25:47 » |
|
The 4th game of the Deep Junior - Kasparov match ended in a draw. This time, Kasparov played extremely cautiously. The score is now 2-2, with 2 more games left. In the next one, Kasparov is going to play white.
Here is the 4th game:
http://www.x3dworld.com/Entertainment/chessMVM/DGT/Game4/Game4.html
Some comments from chessbase.com:
http://www.chessbase.com/newsdetail.asp?newsid=771
Junior Shows What It Knows in Game Four
<snip>
In the maneuvering session that followed, Junior only occasionally exhibited the lack of comprehension of the position that you might expect. The board was full of pawns and the position required artifice and planning, not brute force and tactical wizardry. In the olden days this would have been dark clouds on the horizon for the computer player.
<snip>
Computers can't plan, of course. They must see every move in sequence, beginning from the position on the board. They cannot visualize strategic goals or fantasize about how to achieve them.
A human can look at a position and think, "Wouldn't it be nice if I could get my knight to the b4 square? It would be very strong there, how can I make that happen?" A computer thinks, "I go there he goes there I go there he goes there I go there he goes there I go there and my evaluation goes up to +0.21." If that analysis reaches as far as achieving the human's visualization, the same end can occur by very different means.
Top programs like Junior can perform that analysis so fast and evaluate the results so well that they can simulate human planning. This is really put to the test in closed positions in which human visualization is superior to computer calculation. In game four Junior showed us that this gap is getting smaller all the time.
<snip>
When a program evaluates a position almost all of the weight comes from material balance. "Who has more stuff?" the computer asks after calculating each move. This sounds a little brain-dead, but it asks this question a few million times per second! Along with that fundamental calculation, programs add in pieces of knowledge, the sort of thing humans use: space, king safety, passed pawns, etc.
So when a program plays a move that involves giving up material for positional considerations it is betraying its very nature. Junior does this so regularly, and so well, that game after game you can see that co-programmer Shay Bushinsky said quite a lot when he said in a press conference this week that the difference between Junior and other top programs was that Junior gave less importance to material.
Before we go too far we should point out that just because a move isn't computer-like doesn't mean it's the best move! But Junior is consistently good at justifying these curious decisions. Its 24.a5 in game four was just such a move. The commentators soon turned against Junior's aggressive decision, saying that the passed b-pawn would eventually be surrounded and chopped down, leaving Black with the advantage. 35 moves later when the draw was agreed the pawn still stood on b6 like a statue in the Louvre that had seen the French Revolution take place around it. If Junior wasn't completely write, it wasn't wrong either.
<snip>
Junior showed that it could make something from nothing instead of foundering when there were no targets to aim at. It made a few pointless pawn moves only to follow them up with useful ones. Most importantly, it showed that Kasparov couldn't just set up a defensive wall and take a nap. Junior pressed him for almost six hours, although the last 30 minutes were purely academic.
|
|
|
|
|
rhinoceros
Archon
Gender:
Posts: 1318
Reputation: 8.06
Rate rhinoceros
My point is ...
|
|
Re:Man vs Machine: Kasparov - Deep Junior
« Reply #5 on: 2003-02-07 14:30:41 » |
|
Kasparov - Deep Junior 2.5 - 2.5
Deep Junior surprised everyone by trying an unexpected (for a computer) piece sacrifice with no computable justification. Kasparov defended cautiously and the game ended quickly in a draw. Only one game is left (to be played today, Friday), with Kasparov playing black.
Here is the 5th game:
http://www.chessbase.com/games/2003/x3d5.htm
Comments from chessbase.com:
http://www.chessbase.com/newsdetail.asp?newsid=777
Kasparov-Deep Junior draw after stunning sacrifice
Garry Kasparov was determined to win his last white game against the computer. But on move ten Deep Junior produced a stunning piece sacrifice that left its opponent reeling. It's unclear if the combination was sound, but Kasparov was not going to risk testing it over the board. He quickly forced a draw by repetition to keep the score level at 2.5:2.5. Read our illustrated report.
HAL or Tal? Junior Stuns Kasparov in Game Five
Game five of the Kasparov-Deep Junior match was the shortest game so far, just 19 moves. It ended in a draw after Junior played an amazing bishop sacrifice on move 10 that led the game to a perpetual check draw. A stunned Kasparov found the best moves to survive the black attack and declined to play a risky attempt to continue the game on move 16. The match is tied 2.5-2.5, setting up a high-stakes battle in Friday's game six.
When Deep Junior played 10...Bxh2+ it smashed Garry Kasparov's kingside and many of our conceptions about computer chess with a single spectacular move. I haven't heard a chess audience make as much noise since the Great Pea Soup Riots of Wijk aan Zee '94.
If we needed any further proof that Deep Junior is something beyond what we have come to think of as computer chess, game five provided it. From a quiet opening in a standard-looking position, the Israeli brainchild of Amir Ban and Shay Bushinsky sacrificed a bishop out of a clear blue sky to drag Kasparov's king out into the open. Attacking with only a queen and knight, Junior put the white king under heavy fire.
When a machine plays something like 10...Bxh2+ against you, one thought goes through your mind. Unfortunately this is a family website and I can't print that thought here. Your second thought is, "So should I just resign?" You could tell from Kasparov's always-expressive face that he, like everyone in the room, was having painful flashbacks to game six against Deep Blue in 1997, when he was blown off the board after a piece sacrifice by the machine.
Kasparov collected himself, realized he wasn't being mated, and played the next few forced moves quickly. He captured the bishop and played his king from the frying pan into the fire on g3. Junior feinted with its queen, keeping it close to Kasparov's king. The climax was reached on move 16, when it was becoming clear that Kasparov could allow a perpetual check draw or play a dangerous line and continue the game with his king in the center.
As the strongest human chessplayer in the world sank into thought, the commentators and our computers were frantically analyzing the consequences of his options. 16.Bxh7+ and 16.Ng3 both allowed Black to force a repetition. 16.g3 was the only move to continue the game. Junior would have many attacking options, and could Kasparov be computer-perfect in his defense? Might the position be lost anyway?
Those questions will only be answered in the notes. Kasparov thought for over half an hour and allowed the perpetual check after 16.Bxh7+. The game was agreed drawn on move 19. A short draw with black against Kasparov cannot be called anything other than a tremendous success for Junior, and the spectacular method in which it was achieved is an extra feather in Junior's yarmulke.
Although most programs we've checked with at first say that White had good chances after 16.g3, it was not a move for a human to play against a computer. Black gets good compensation for the piece both in pawns and/or in attacking chances. How much Junior actually saw and how much Junior estimated based on subjective factors we do not yet know.
<snip>
This sets up a game six showdown that also echoes 1997. The match is tied and Kasparov has black in the final game. Could this deja vu be what Ban and Bushinsky had in mind when they chose to have black in game one and white in the final game?
Also: A Fark photoshop contest:
http://cgi.fark.com/cgi/fark/comments-voteresults.pl?424899
|
|
|
|
|
|
Logged
