Wednesday, 24 January 2018

Trying out Stockfish 8

Having used a combination of old versions of Fritz and Rybka to error-check my chess analysis for years, I decided to give Stockfish 8 a go.  Ironically, this was prompted by Stockfish's much-publicised losses against Google's AlphaZero.  Yes, even bad publicity can be good.

I'm already finding that Stockfish sometimes disagrees with what Fritz and Rybka had to say on particular opening lines, and more often than not it tends to be right when I examine its suggestions closely.  For instance, I've previously examined the 13.Nxh7 sacrifice in Thiele-van Perlo, corr. 1987 in the Göring Gambit in the following position:

My previous examinations of the position suggested that 13.Nxh7 was dubious, and that 13.Ne6 was better, but a further examination with Stockfish suggests that 13.Ne6 is dubious because of 13...Bxe6 14.fxe6 and now the computer inconveniently points out 14...Ng8!, after which I can't see how White makes further progress.  Meanwhile, 13.Nxh7 appears to be sound, and may well be the best move in the position.  Thiele rather erred after 13...Kxh7 14.Bh5 g5 15.fxg6+, when after 15...Kg7 the g6-pawn blocks White's avenues of attack.  After instead 15.h4, it appears that White has at least enough for the piece in all lines.

I decided to feed Stockfish 8 a position that David Norwood used to show computers back in the 1990s, and which is sometimes thus known as the Norwood Position:

 Highlighting how far computers have advanced in the past 20 years, Stockfish 8 doesn't even look at the rook on a5, recommending that White shuffle the king, although it does erroneously assess the position at -17 pawns in Black's favour (the correct assessment is that it's a draw!).  If you enter 1.bxa5, it immediately gives mate for Black in circa 18 moves.

More challenging for the computer is if you replace the b-pawn with a bishop:

No, this wasn't originally my idea.  I can't recall where I first read it, but it is discussed in the Computer Chess News Sheet June-July 1994, so this might well be where the revised position originates from.  I recall feeding it to Fritz and Rybka some time ago and both computers insisted on grabbing the rook, but it might have changed with the latest commercial versions.

Stockfish 8 recommends 1.Bxa5 for a couple of minutes, assessing it as -5.2 pawns in Black's favour, and gives 1.Bb4 (the correct move) as -9.5 pawns in Black's favour, but then it picks up on 1.Bxa5 b4!, and within another couple of minutes it rejects 1.Bxa5 and gives 1.Bb4 followed by shuffling the king as best.  So even the revised version no longer stumps today's strongest computers.

There are of course still blockade positions that are even beyond Stockfish 8, but as computer AI continues to improve, they have to be more and more inventive.  I'm left wondering how AlphaZero, with its Monte Carlo method of calculation, would fare.  We might never know!


  1. My current engines are (in order of preference) Houdini 4, Deep Rybka 5, and Deep Fritz 13. But really I think any recent engine is good – once you've got used to it and begun to appreciate the instances when you can't quite trust it.

  2. Since 1st of February there is Stockfish 9 available.
    In my article I wrote that the top-engines gain averagely about 52 points per year in the last 2 decades. You can do the math to know how big the gap is between your old and new engines. It is no surprise that many of your old evaluations will change.

  3. Yes, I've now added Stockfish 9 to my engine stack as well. It's gone in at No.2 in preference.