21 February 2010

A Few Novel Ideas

While researching yesterday's post on Kasparov and chess960 -- Why Not Announce Positions Beforehand? -- I encountered a few novel ideas related to chess960.
  • Kasparov on chess 960 and preparation (chess.com): Re Kasparov's emphasis on retaining the possibility of opening preparation, 'Where's the fun in playing an opponent who spent the last month analyzing some opening sidelines with Fritz/Rybka? Is chess just about rewarding hard work?'

Once again, the dichotomy between chess at the professional level and chess at the amateur level enters the discussion. I never get tired of saying, 'the first rule of chess is to have fun', but perhaps this doesn't apply to professionals.

  • Kasparov on Chess960, et al. (iccf.com/forum): 'For serious correspondence chess, as opposed to casual correspondence chess, playing chess960 games is a step in the WRONG direction. The reason is simple : the human knowledge effect in the games will be further reduced since the engines that are already affecting classic correspondence chess have zero problems adapting to chess960.' • Followup: Chess960 the answer for serious CC?.

Since Kasparov said nothing about correspondence chess, this is a non-sequitur, but nevertheless an interesting opinion. I'm not aware that anyone has suggested that chess960 changes the impact of engines in correspondence chess and I suspect that engines have the same advantages (and disadvantages) in both traditional chess and in chess960. This is certainly true after the opening, when the games become indistinguishable. As for the opening, what difference does it make if the correspondence player uses an engine on the first move in chess960 or on the Nth move in traditional chess when theory is coming to an end?


HarryO said...

Hi there
I have tried some experiments with 960 chess engines. I used Deep Rybka 4 960 on a dual core. I can absolutely confirm beyond doubt that the modern chess engine DOES have a lot of trouble adapting to the chess960 opening. Rybka's opening priority game after game seems to be:
1) Castle first if it is possible
2) Move out the knights
3) Move the knights twice if there is a concrete attack or an outpost that has no short term counter.
4) Attack as soon as possible with the queen
5) Let the bishops sort themselves out over time.

In other words, the best chess960 engine in the world at the moment is beatable by a 2000+ chess960 rated human on classical time controls more often than not on an average PC. I do not know what happens when you increase the number or cores, but suspect that the search horizon will still be inadequate for the machine to find the best plans. The only reason a 1900 player looses to Rybka is because of blundering. The machine cannot be trusted on it's opening choices and even in many 960 mid games, will seriously mis-evaluates positions where it's pieces are still on the board but hopelessly trapped (effectively useless long term).

As far as "Advanced Chess" 960 computer assistance and online correspondence chess, I see no issue at all if computers are used (so long as it is declared), because human skill is still the biggest single factor in deciding a 960 game in this era. All that has to happen is that these types of games have to be declared as computer assisted.

In order for 960 engines to play the opening better, some serious attention will have to be given to force feeding the machine a general set of reasonable rules to follow that work in most 960 openings. The machine does not have a search horizon of great enough depth to understand how to find the best piece setups and I doubt it will for some time until new technologies emerge.


HarryO said...

I do have to be very careful! I mean no disrespect to the programmers of chess engines such as Rybka! I have to be careful because the computer's choice of opening is highly dependent on it's search depth. Within that window, it will evaluate many moves, but essentially the final choice is dictated by noise (where the evaluations between moves choices are so close and dependent on the length of time the machine has calculated). So the list of rules that I gave above is not actually true in reality!. It is just a trend I have noticed when the engine is forced to make a quick decision (because the tester is impatient!)

Where the machine really falls down is in gambit lines. These are lines where a pawn is sacrificed for positional gains beyond the machine's search depth. Here is a great example.


If you plug this into Deep Rybka 4 960, there are a lot of really closely evaluated moves that Rybka could choose. However always the machine misses a possible gambit line that I think a human would like to play as white:

1. d4....d5
2. c4!?...dxc4?!

The machine will not play c4!? no matter how long you give it! Does the machine know something I don't perhaps :-) Please anybody take a look at this gambit and show me that it is definitely not a good move for white! From my own experience of playing this position, I have enjoyed playing the gambit, but the machine evaluates it badly until eventually it's evaluations all swing positive and the gambit does seem to work.

I stand by my claim that a +2000 rated chess960 player would beat the best machines of today quite regularly, even if that claim is untested.