Quality of Chess Games
What's the best tournament chess game of all time? If by 'best' one means 'best played' then I'm afraid the answer is Zappa vs. Fruit. In this most recent world computer tournament, Zappa scored an astounding 10.5 out of 11, a better performance than any human has ever had in a human world championship, and against a stronger field than any human world champion has ever faced. Fruit came in a clear second, so this is the only tournament game we have between the two strongest chess players ever created. Of course, you'll soon be able to buy the commercial version of Zappa and have it play against itself, resulting in a string of games most of which are better than any game ever played between two humans. Welcome to chess in the 21st century.

Some humorous notes: Zappa and Fruit were both written by lone grad students in under two years. Dark horses obliterating the field is a common thing in AI. Zappa's lone draw was ironically against the program which lost every other game in the entire tournament.

Zappa's high win rate as black against extraordinarily strong opposition indicates that the frequent grandmaster claim that they can automatically draw as white is a bunch of bullshit. In a match between Zappa and any human on the planet with Zappa taking black every game and an even score giving the match to the human, I'd put my money on Zappa in a heartbeat.

Now that computers are clearly better than humans at chess, the question arises, can computers attempt to guess the strength of a game's play based on the moves in that game? And can we use that method to evaluate 'classic' games? Do we really want to?

As for rating play based on moves, a simple algorithm which compares the strength of the move actually played to the strength of the best move will probably do nicely. It's necessary to have some formula for determining how much a blunder counts against you versus a series of small mistakes, but that should be straightforward to determine with some experimentation. There's a bit of a problem with endings which the computer totally works out, since then the difference between the best move and the second best is likely to become very large based on what the computer happens to have worked out completely, but it shouldn't be terribly difficult to deal with that issue either.

On the question of whether classic human games are any good, I'm afraid the answer is most likely no. There is nothing special about grandmaster human strength in the continuum from random move play to perfect play, and the computers are now better, rendering human grandmaster games just as childish to computers as human international master games are to human grandmasters.

That, by the way, is why computers these days typically throw out the opening book after the first few moves, if not ignore it altogether. The opening book was generated by humans whose skill level was far worse than the computer, and the only advantage they have is the sheer number of hours put into study of the opening position. You don't have to get many moves deep before that advantage goes away.

However, there computer evaluation of human play may give an opportunity to find the occasional game in which the humans understood something which the computer still doesn't grok. In each game, the computer will evaluate one side as having played superiorly to the other. We could have the computer evaluate all available grandmaster games, and find ones in which the computer rates the play of the side which lost as superior to the side which won. The largest gaps in that direction will undoubtedly be games in which someone resigned a won position, but there are probably many games not too far down the list which indicate something useful.