Can Houdini 3 be beaten?

While many people regard chess supercomputers as unbeatable, I've managed to defeat the program in three separate variations of the Two Knights' Defense in the past week. Here is a real crush:

If the position is simply winning for one side, as here, then it shouldn’t matter who the players are i.e. if it is a win, it is a win, Caissa doesn’t care about ratings, it is only the position on the board that matters.

Moody (1700)-Houdini 3 (3334)

1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 d5 5.exd5 Nd4 6.c3 b5 7.Bf1 Nxd5 8.Ne4 Qh4 (Dr. Hans Berliner studied this intriguing gambit for decades and even used the strongest computer available at the time, HiTech to validate his theories. The gambit as I and others have demonstrated is unsound.

9.Ng4 Bg4

10.f3 e4 (This move should raise warning flags; in an ultrasharp position where every tempo counts, being forced to make a pawn move can’t be a good idea)

11.cxd4 Bd6

12.Qe2 (Bxb5ch is also winning but World Champion Estrin got lost in the position (thankfully!) and Berliner played one of the greatest endgames of all times. Qe2 is an idea of the ICM Walter Muir) Be6 (Repositioning a developed piece can’t be good. O-O is only marginally better; after 12…O-O 13.fxg4 Bxg3ch 14.Kd1 & any decent chess engine will crunch the win for White, but Be6 is geared to the superficial idea 13.fxe4? when the Black Knight is driven to the strong square b4 with compensation. Here White has three winning plans: 13.Qf2 holding the piece, 13.Qxb5ch forcing the King to move and just winning a free pawn, and 13.Nc3!

With a single move White offers a Rook to squelch the attack, knowing he will have three pawns for the exchange and the better position i.e. he is simply winning)

13.Nc3  Nxc3

14.dxc3 Bxg3ch

15.hxg3 Qxh1

16.Qxb5ch Kf8

17.d5! (An idea of IM Jeremy Silman. He stated, “This position sucks for Black”) Bf5

18.fxe4 Bg4 (Bxe4?? 19.Qb4ch++-)

19.Qc5ch Kg8

20.Qxc7 h5

21.Be3 h4

22.Qf4 Bd7

23.O-O-O The win is just a matter of technique even for a Class B player like me. All I have to do is push the pawns.Qh2

24.Kb1! It is small moves like this that constitute “technique” Ba4 The horizon effect. This simple failing of computers programmers no doubt thought was a thing of the past.

25.b3 Bd7

26.c4 (Charge!) hxg3

27.Re1! (Eternal vigilance is required against computers. Here I am concerned about ideas like Rh4/Bg4) Rh5

28.Qc7 Bg4

29.c5 a5

30.c6 Qh4 (writhing)

31.Qf4 a4

32.b4 Rd8

33.b5 Qe7

34.Bd2 Bc8

35.b6 Re8

36.d6 Qd8

37.c7 Qd7

38.Bc4 Rc5

39.Rc1 Rxc4

40.Rxc4 f6

41.Qxg3 Re5

42.Qd3 Kf7

43.Rb4 1-0

While supercomputers are dangerous opponents, they can be beaten.

sloughterchess!  It's been so long.  I didn't know you were still on the site :)

what was the time control setting for Houdini 3?

While it was game in 60, Houdini tends to ignore time limits and typically took several minutes/move.

This game challenges the notion that supercomputers can’t be beaten:

1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 d5 5.exd5 b5?! (The Ulvestad Defense)

6.Bf1 h6? (book)

7.Nxf7! (With this obvious move White has a clear, close to winning advantage) Kxf7

8.dxc6 (Now, 8…Bc5 is supposed to give good play, but this is an illusion. Houdini evaluates this position as +- i.e. a “book” move leads to a dead lost position. This is far more common than you might believe; the moral of the story is---never let a machine play book without “fact checking” the variation i.e. a cursory search for gross blunders like this variation) a6 (Houdini is out to a depth of 22 here and sees nothing better)

9.a4 (Until this game I had always played Be2 here) Bc5 (Black has made progress; now the White advantage is “only” +/-)

10.Be2 Ne4

11.O-O Nxf2

12.d4 (Of course---this is why Houdini, left to its own devices, would have seen this and avoided it with 6…Nd4 transposing to the Fritz---which is “only” +/- Black would love to play a discovered check hereabouts but the Knight is pinned) Bxd4

13.Rxf2 Bxf2ch

14.Kxf2 (It is obvious that the “attack” is over and Black is struggling to survive. What should be clear is that if White had good endgame skills, this is very close to a forced win for White. When you consider that White has a material advantage and a huge positional edge due to the severely compromised Black pawn structure, White should win this. This assessment may seem harsh, but did you notice that Black hasn’t a single developed piece except for his King?! Depth = 24) Qxd1?!

15.Bxd1 Bf5

16.Nc3 (This is the second time I reached this position. On move 17 in the first game I opted for a “safe” option and only drew. At the highest levels it is not acceptable to play safe when you have a significant advantage. It is just as bad to draw a clearly favorable position as it is to lose an equal position. Where it is even worse are those tournaments where a win is awarded three points. It is pointless to play safely when wins are so critical so I resolved to “crunch” the position) b4

17.Nd5! Rad8 18.Bf3! e4 19.Bh5ch g6 20.Ne3! Rd6 21.Nxf5 Rf6 22.Bxg6ch Kxg6

23.g4 h5 24.h3 hxg4

25.hxg4 (Some times in the course of a game it is easy to lose track of the fundamentals; here, for instance, we are in endgame where White has an active King and a significant material edge that can be converted to a passed pawn on the Queenside) Kf7 26.Bd2 Rxc6 27.Rc1 b3 28.cxb3 Rh2ch 29.Ke3 Rxc1 30.Bxc1 Rh3ch 31.Kxe4 Rxb3 32.Kd5 Rb5 33.Kc6 Rxa5 34.Bd2 Ra2 35.Bc3 a5 36.Kxc7 Kg6 37.Nd6 a4 38.Nc4 Kg5 39.Kb6 Kxg4 40.Kb5 Kf5 41.Na3 Ke4 42.Kxa4 Kd3 43.Kb5 Ke4 44.Kc4 Ra1 45.Nb5 Rd1

46.b4 Kf5 47.Nd4ch Ke5 48.Kc5 Ke4 49.Nc6 Kd3 50.Bb2 Rb1 51.Ne5ch Kc2 52.Nc4 Rh1 53.b5 Rh5ch 54.Be5 Rh4 55.b6 Kb3 56.Nd2ch Ka4 57.b7 Rb4 58.b8Q Rxb8

59.Bxb8

I resigned for Black here. It is a book win, but difficult to implement in real life. Even GM’s sometimes draw here with White. In this situation Houdini had overstepped the time limit about a half an hour earlier (Game in 60); besides, I knew that I couldn’t win the endgame!

Here is a simple win in the Ulvestad. Once again Houdini never comes close to equalizing after the blunder 6...h6.

The questions being addressed are: 1) “Can Houdini 3 be beaten if it plays the Two Knights’ Defense?” 2) “Are all variations of the Two Knights’ Defense better for White?” What I hope to demonstrate is that the Two Knights’ Defense is so good for White that even a super computer can’t equalize against best play by White. One of the practical results I have discovered is that in the middle game, if the evaluation ever rises above +1.2, then it is a win for White.

What I intend to do is to test every major variation of the Two Knights’ Defense under tournament conditions. I will test under what many players would consider valid i.e. no sight of either the moves or evaluation of the computer as it moves under supervision by a tournament director. Due to time limitations I will play the games at a time limit of 2 hours for the entire game with a 10 second bonus for each move; Houdini will not be allowed to lose on time.

This is an unsupervised game but shows the problems facing Black in the Ulvestad:

1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 d5 5.exd5 b5?! (The Ulvestad)

6.Bf1 h6? (Perfectly acceptable according to MCO 15; according to theory Black gets a good game with an early Bc5; as demonstrated here, h6 is a very weak move. Black should play 6…Nd4 and transpose to the Fritz as Berliner did in his classic game against Estrin for the Correspondence World Championship)

7.Nxf7! Kxf7

8.dxc6 a6?!

9.a4 Bc5

10.Be2 Ne4

11.O-O Nxf2

12.d4 Bxd4

13.Rxf2 Bxf2ch

14.Kxf2 Qxd1

15.Bxd1 Bf5

16.Be3 (Houdini prefers Bd2; this is more active) Be4

17.Nc3 Bxc6

18.Be2 (I don’t see Houdini’s point of Bh5ch; I intend to control all invasion routes by the Rooks to the 1^st, 2^nd, and 3^rd) Rhf8

19.axb5 Kg8ch

20.Kg1 axb5

21.Rxa8 Rxa8

22.Nxb5 (According to my theories White should not even consider Bxb5. The reason is that Nxb5 gains a tempo over Bxb5 i.e. the Bishop can access b5 in one move, thus the move sequence Be2/Bxb5 does not gain time. Since the Knight cannot access b5 in under two moves, Nxb5 is a developing move. While it might not seem so, this is still a time/tempo attack even though it is an endgame.) Rb8?!

23.c4! (I welcome an endgame of two Bishops and pawns versus Rook and pawns. Again, according to my theories c4 gains time) Bxb5

24.cxb5 (Black has a dismal defensive task here) Ra8

25.Bd2 (I regard this as a simple technical win. After Bc3 all the White pawns are defended and it is just a matter of time before White creates a passed pawn on the Queenside) Kf7

26.Kf2 Ke6

27.Ke3 Ra4

28.Bc3 (According to my theories this is still a time/tempo attack i.e. all White’s moves are developing moves) Kd5

29.Bf3ch (Not to beat a dead horse, but this is a developing move gaining time with an attack on the King. Up to this point I have only used a few seconds/move because I am just taking a quick look at tactical shots 1-2 moves ahead and just developing as quickly as possible. It is only after this last move do I start actually “crunching” the position; the previous moves were just following my theories. I know I am winning but don’t have a clue how to win it so I just decide to make a few harmless moves until a plan appears.) Kd6

30.Be4 Ke6

31.h3 Kd6

32.g3 Ra1

33.Bc6 Ke6

34.Ke4 Rd1

35.Bb7 g6

36.Bc8ch Kd6

37.Bg4 Rb1

38.Bxe5ch Kc5

39.Be2 c6? (Houdini was out to a depth of 28 and could see nothing better; now I have a plan---create passed pawns on both wings)

40.bxc6 Kxc6

41.h4 Kd7

42.Bc4 Re1ch

43.Kf4 h5

44.Bg7 Ke8

45.Bf6 Rg1

46.Be5 Re1

47.Bb8 Kf8

48.Bd6ch Kg7

49.b4 Rd1

50.Be5ch Kf8

51.Kg5 Ke7

52.Kxg6 Re1

53.Bf6ch 1-0

White followed a “cook book” recipe out to move 28 and then just had to win a won game:

Is this guy a scammer, a joker, or simply delusional? I only went over the first game he posted, but it was clear to me that he wasn't playing a full-strength Houdini of any version. Wassup, guys?

4.Ng5? moving the same piece twice.  Either castle or move another piece. 

Actually an acceptable book move. This move wasn't one of my concerns. (Houdini was playing Black, anyway.)

I only had to go through the first 20 moves or so to know you weren't playing a full strength Houdini, or even Houdini at all. A few of the moves played weren't even in Houdini's top 5... for example 10. e4?? is a blunder according to Houdini.

Yeah, I think we do understand. You're substituting moves for Houdini that Houdini would never play, and then you're claiming that you beat Houdini. Uh, OK...

Maybe we could set up a match between you and Houdini 3 through the community...this will lay this to rest...

If a willing participant would come forth that owns Houdini 3...the two of you could continue to exchange moves in the forum and we could follow along with updated game sequence boards.... ?

Any takers, have this already been proven this is a sham ? Let's find out...

At what point in the first game does your cherry picking of your opening theory end and Houdini is allowed to play for itself? Some of the more blatant poor moves that Houdini would never play are 9, 10, 12, 24, 27, 32, and 39. So your opening theory goes to move 39? Impressive! (I seriously think you should check your Houdini configuration settings to see if it's really playing at full strength.)

Probably this thread should be named "Can sloughterchess-assisted-and-manipulated Houdini 3 be beaten?"

By the way I beat Houdini few times a day while training theoretically won endgames against it

Probably the only way most humans can and should be able to beat Houdini. :)

?

Castling would be a blunder, not 4. Ng5, a book line which wins a pawn by force, for some positional compensation.

"...Houdini tends to ignore time limits..."?  That should have been your first clue that something ain't right.

ScorpionPackAttack's comment is a nice example of how general principles can be easily abused if they are given priority above specifics of concrete position.