That really is an excellent point.
My feeling from several of the published games was that AlphaZero was stunning at weighing complex positional factors (some of them at least partially comprehensible to most chessplayers!) against material, and Stockfish was found wanting at getting the correct balance, with an over-materialistic tendency. However, the sample is so small it can only tell a little of the story.
@JunyJunebug Oh no! I see Skynet firing some missiles right now.....oops...
The original question in this thread was: "Are these the highest quality decisive chess games ever played to date?" The development of AlphaZero may make us change what we think of as a high quality game. For those of us with low ratings compared to traditional engines (aka mostly everybody) all the engine games are high quality chess. I think you have to be exceptionally good at chess to think about which engine games are better/more beautiful/deeper/etc. than others because for me every engine game is such high quality that I can't tell which is superior. If you showed me 10 games where Houdini or Komodo beat Stockfish - I'd say they are amazing too. However, I'm getting the sense that the AlphaZero games make somewhat more sense to us humans than the traditional engine games because of the "learning" nature of AlphaZero. So does that factor make us think of those games as higher quality? What about the drawn games (not that we have any of those for AlphaZero shared with us); when both sides hold their own, does that mean it is high quality?
Of course we do not know what the other 90 games look like! Most of these were draws, but quite a few were AlphaZero wins. It may be they have omitted some totally incomprehensible games. Or maybe AlphaZero never plays like that. Only the DeepMind team know!
Poor Stockfish, just 1GB hash, no opening book and no endgame tablebases (as said by a guy on the chessbase page who analyzed the games). With the endgame tablebases several losses wouldn't have occured...
I am no expert in neural networks so can anybody explain how the neural network can be trained like they said:
" Training proceeded for 700,000 steps (mini-batches of size 4,096) starting from randomly initialised parameters, using 5,000 first-generation TPUs to generate self-play games and 64 second-generation
TPUs to train the neural networks."
taken from their paper and then transferred into a single machine with only 4 TPUs ?
Bear in mind the first generation TPUs were far smaller and slower. The second generation ones are much more powerful. But I am not absolutely sure what stage of integration they are referring to when they refer to 4 units.
I know, but what I don't get is how to transfer a trained network of 64 TPUs to 4 TPUs...
The model is not in the TPUs: they are like CPU or GPU for parallel processing. The model is a data structure with thousands of nodes and millions of parameters that resides in RAM when being used. It is made of tensorflow objects which are just matrices really.
When training they used minibatches of 4096, and each of these samples could be calculated on a separate TPU, I believe.
Certainly there is value in parallel processing when playing as well: it should allow you to generate many MCTS games simultaneously, but there is a greater need for central co-ordination than when training with independent sample self-play games.
Where are the games where AlphaZero played English Opening? I read that it began to play 1 c4 over time
The "opening book" argument doesnt have any sense. Just look at the games, Stockfish came out of the openings well in every game, close to equality in each one, it was simply crushed in the middlegames.