Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model

Re:Mastering Atari, Go, Chess and Shogi by Planning with a L... by bonzini on Thursday November 21, 2019 @02:55AM (#21592215)

AlphaZero evaluates comparably less positions per second than Deep Blue. It doesn't even reach 100k nodes per seconds of I remember correctly, while Deep Blue was in the millions[1]. Therefore, even though the training is done by brute force, the evaluation is way less "brute forcing" in AlphaZero than in Deep Blue.

Not just Stockfish on modern hardware can evaluate fewer nodes than Deep Blue and still beat it left and right; in 1995 Fritz on a Pentium was able to beat Deep Thought II at the world chess championship. Deep Blue and its ancestors, with their custom hardware, were perhaps the "most brute force" of all chess engines.

[1] https://www.stmintz.com/ccc/index.php?id=91692

[ Parent | on HN ]

Re:Mastering Atari, Go, Chess and Shogi by Planning with a L... by mindgam3 on Thursday November 21, 2019 @03:03AM (#21592243)

As you yourself point out, the brute force here refers to solving a problem by throwing more hardware at it.

Number of nodes searched is not the key metric for gauging how “smart” the algorithm is. You have less nodes searched but you only got there by having way more upfront processing.

[ Parent | on HN ]

Re:Mastering Atari, Go, Chess and Shogi by Planning with a L... by lordgrenville on Thursday November 21, 2019 @05:22AM (#21592827)

But that processing happens just once, and then you amortize it over the software's lifetime. Play a million games and you probably come out ahead.

[ Parent | on HN ]

Re:Mastering Atari, Go, Chess and Shogi by Planning with a L... by red75prime on Thursday November 21, 2019 @05:55AM (#21592982)

And what is an estimation of the minimal hardware/time requirements for learning to beat humans at chess?

We need some baseline to call it "brute force".

[ Parent | on HN ]

Re:Mastering Atari, Go, Chess and Shogi by Planning with a L... by anotherone16 on Thursday November 21, 2019 @11:01AM (#21595492)

Seems like ML/AI is still looking for its quantum supremacy moment.

[ Parent | on HN ]