|
Alpha Zero vs Stockfish
AI just shit the bed.
https://www.sciencealert.com/it-took...mind-alphazero
Stockfish is, sorry was, the leading chess computer in the world. It has lots of human input, such as an opening database, strategies, and numerical values assigned to pieces. While Stockfish does recognise that these values are not static, it's weakness appears to be that it does not adjust these values correctly to suit the position.
Stockfish processes 70,000,000 positions a second.
Alpha Zero is the games playing baby of Google's Deep Mind. It's just crushed a professional Go player, a world's first, and it was then told the rules of chess and nothing more. No openings, no piece values, just the rules. It began playing itself, and four hours later begain playing Stockfish in a series of 100 games.
Alpha Zero processes 80,000 a second, while basing its strategy on memory.
As white, Alpha Zero won 25 and drew 25.
As black, Alpha Zero won 3 and drew 47.
Total score - AZ won 28, drew 72 and lost none.
A common theme in the games AZ won was the severe limiting of the mobility of one of Stockfish's bishops. It's like AZ figured out Stockfish's weakness, in that it overvalued immobile pieces.
Another interesting thing that happened is a combination of moves that resulted in a repeated position. The second time the position arose, AZ opted for a different move, which begs the question, why didn't it play that move the last time?
This isn't just huge for chess. This is huge for AI. Four hours it took for a completely different method of analysing chess to shit all over the previous best method of brute force. SF is processing nearly 100x more positions that AZ, but AZ has a superior understanding of the value of pieces based on their mobility.
AZ is processing and learning, while SF is merely processing.
Not sure if exciting or scary, but I have a boner.
|