The game will go down in history as one of the great queen sacrifices, and fans were given an absolute treat. Let’s skip straight away to the key points, always enjoyable to see again and remember.
Above is the position after 18…e5 and both players had been playing fairly quickly, which led one to understand that both had seen this position and line in their preparation. White played here the beautiful queen sacrifice 19.Qxf6!! Here is where it makes sense to see what the engines saw since no one would enter this position nowadays at such a high level in the opening without having studied it thoroughly. All engines see the queen sacrifice almost instantly, but the main line they display involves recovering the queen with a roughly equal position. After the more or less forced 19… Bxf6 20.Nd5 Qd8 we reach the next key crossroads.
Presuming no one is privy to exceptional hardware such as 16 cores or more, a look at Stockfish 8 or Houdini 5 for at least three minutes on a good quad-core (an Intel i5 at 4 x 3.0GHz for the purposes of this article) will show a roughly equal evaluation with the main line 21.Nc6 (the black queen is trapped) 21…Bxg4 22.Nxd8 Bxd8.
Caruana's sparkling win over Nakamura was the highlight of the London Classic (photo by Lennart Ootes)
Instead, Caruana uncorked 21.Nf5!!, leaving the material imbalance on the board with White down a queen for two pieces. Thus when Polish grandmaster Krasenkow commented in his Round Six analysis that engines struggle with such positions, and that it could only be the result of human ingenuity and intuition, it was no doubt after seeing no engine coming up with it in a reasonable time.
"I must say that such "real" queen sacrifices are an Achilles' heel of analytic engines. They almost always evaluate positions with, say, two minor pieces for a queen in favour of the strongest piece, while in reality, in human play, there may be a more than sufficient compensation for the material. Therefore such positions should be analysed and evaluated "manually", and kudos to Caruana and his team for doing that!"
There seems little doubt this is what happened to Nakamura and his second(s), who used an engine, but did not insist after seeing nothing after a few minutes. How long would they take on a normal quad-core?
Stockfish 8, a top free engine, using 2GB hash tables, takes around eight full minutes. So the user would need to wait a full eight minutes before the engine announced it had made a breakthrough.
The new member of the family of elites, Houdini 5, takes about half that time, and finds it in 2 minute 47 seconds. By no means is this to suggest Houdini 5 will always outdo its rival by a factor of three, since solving a position varies enormously from one to the next. The point is that it too takes nearly several minutes to find Caruana's 21.Nf5
Still, readers may recall an article that was posted a few days ago describing the special Tactical Mode in Houdini 5. An option in the engine that turns it into a single-minded combination seeker, a bit like the late great Tal, hell bent on trying to find a tactical shot.
When Houdini 5’s Tactical Mode is enabled it finds Caruana’s 21.Nf5!! In exactly… zero seconds! Not only that, but it finds the full winning continuation, and displays a large plus score for White.
So what if you analyzed the position before the queen sacrifice? How long would it take to see 21…Nf5 in its main line? Meaning how long would it take to see the complete winning sequence from 19.Qxf6 until 21…Nf5? The answer is two seconds!
Once more, this emphasizes the usefulness of the Tactical Mode in Houdini 5 as a special tool to help analyze just such chaotic positions, helping avoid fatal oversights.
It should be mentioned that Priyadarshan Banjan, editor of ChessBase India, correctly pointed out in his article on Round Six that as soon as one had opened an engine in ChessBase 13 or 14, one would also see the top three engine analyses from others in the Live Book. Your engine might have only analyzed a second and seen nothing of note, but others might have already seen this in great depth, and that analysis is immediately available.
As you can see above, Houdini 5 (normal mode) has only been running for a few of seconds, but on the right are analyses of Stockfish 8 at a depth of 60 plies (!!). The brown colored lettering means it was analyzed for the equivalent of 60 minutes on an i7 processor. No matter what you are analyzing, never forget to consult the Live Book as it is quite possible someone has already analyzed your position to great depths, saving you the time of doing it yourself.
.
V.Kramnik - P.Leko, 0-1
Match World Chess Championship (MWCChamp) game 8, 2004/Oct/07.
.
.
Next from the above FEN position, Leko moved 25.. Qg6-d3!, a winning move that Kramnik's home computer had misjudged.
.
Kramnik blamed his helper seconds for not adequately testing this Qd3 move, and the seconds agreed to the blame.
.
However, the Fritz (or whatever chess engine) they used needed an exceptionally large amount of time to detect the strength of Qd3. The seconds cannot give so much time to every possible opening move during pre-game at-home preparation.
.
Maybe in the future, most opening preparations will be tested by Houdini in both regular mode and its new tactical mode?
Oh and kudos to fabi and his team for finding such a move - it's incredible that stockfish8 starts to understand that Nf5 might be a good move only around 10 minutes at decent quad analysis (and it takes even more time until it realizes that it's actually winning). Very intruiging position. For example another famous "computer analysis bludner" in the match Kramnik vs. Leko where Leko refuted Kramnik's preparation OTB - at that time there were saying that decent computer would see the hole in their preparation in about 90 seconds (time they couldn't spare at that time - since they had many positions to analyze) - while today Stockfish 8 on decent quad solves it in matter of couple of seconds. Therefore those 10 minutes nedded for Fabi - Naka position seems like eternity from computer chess perspective, which makes this position so much more impressive and strangely enough - from the way fabi has described this position - despite its complexity the position might be actually easier to understand by top players than a top engines.
This is also a Q sac for two pieces and long term positional compensation...Black gets the dark squared B and usually ends up winning at least an exchange afterward (after 13. Rc1 Nxc4 14. Rxc4 bxc4, as the stem game and many subsequent games continued)
The Kings Indian has often been an openings the computers (the less recent computers, at least) seem to misevaluate the most....And this Queen sacrifice is very difficult for strong humans to play against with the white side, which is why many of them decline it and simply recapture the Knight on move 11
Tal said "There are two types of sacrifices - correct ones, and mine"....But this Queen sacrifice of Tal's seems to be correct, at least in human praxis....And Caruana's Queen sacrifice is also correct, apparently
"When Tactical Mode is active, Houdini modifies its search process and uses some clever tricks to focus on tactical solutions in the root position. In Tactical Mode Houdini becomes a very skilled tactical problem solver. ...
Note that the focus on tactical moves will reduce the playing strength of the engine, the Tactical Mode is significantly weaker than the normal playing mode.
One can compare the Tactical Mode to what a human player does when he/she tries to solve a test position in a chess magazine or website. He/she will scan for tactical, forcing moves, threats, sacrifices etc. That way a lot more tactical solutions can be found than what the player would actually see and decide to play over the board, during a normal game. The "this is a test position" message changes the search strategy.
Houdini's Tactical Mode works exactly the same, very human-like: it will spend a lot more time looking for tactical moves, threats etc. In most positions this doesn't provide any benefit - most of the time there is no tactical solution available and the increased focus on tactics is just wasted. But in test positions (selected because there IS a tactical solution) it works extremely well."
Is 'Tactical Mode' mostly telling the engine to Not Prune moves that are probably hopeless by routine standards? By skipping the logic which evaluates moves, those CPU cycles could be used for brute calculation instead?
.
I vaguely recall that older chess engines had "personality" options, including one for emphasizing tactics over positional efforts. Is Houdini's 'Tactical Mode' a better implementation of that basic concept?