Child Prodigy defeats World Champion.
On Thursday, December 7, 2017 at 12:05:48 PM UTC-7, Rainer wrote:
Now that AlphaZero is all the rage, it's time to put some things into
1. AlphaZero ran on superior hardware which should be worth much more
than a +102 Elo performance. So we cannot say that AlphaZero is the
stronger chess engine.
2. The testing conditions are unclear. I have the suspicion that they
were not fair for Stockfish. Hikaru Nakamura also expressed his doubts,
so it isn't just me.
This is indeed quite true.
So it is not accurate to claim that AlphaZero is superior as a chess
program to Stockfish - run a version of that program adapted to use
your computer's hardware, and it would play much worse than Stockfish.
So why is this exciting?
For one thing, Stockfish was evaluating 70 million positions a second,
while AlphaZero was evaluating 80 thousand positions a second.
So AlphaZero somehow obtained a lot more chess knowledge than Stockfish
has. Studying what it came up with as settings for its neural network
would perhaps make it possible to upgrade Stockfish to do what Komodo
is good at: finding good positional moves in quiet positions.
Already, studying Alpha Go has led to improvements in conventional Go-
playing programs - the latest version of CrazyStone gained 100 Elo this
way. This happened by adding "deep learning" techniques to its