What was the average decision speed pf Alpha Zero in the recent Stockfish match?


The match got a lot of press, and I doubt anyone is surprised that Alpha Zero crushed Stockfish.

See: AlphaZero Destroys Stockfish in 100 Game Match

To me, what's really salient is that "much like humans, AlphaZero searches fewer positions that its predecessors. The paper claims that it looks at "only" 80,000 positions per second, compared to Stockfish's 70 million per second."

For those who remember Matthew Lai's GiraffeChess:

However, it is interesting to note that the way computers play chess is very different from how humans play. While both humans and computers search ahead to predict how the game will go on, humans are much more selective in which branches of the game tree to explore. Computers, on the other hand, rely on brute force to explore as many continuations as possible, even ones that will be immediately thrown out by any skilled human. In a sense, the way humans play chess is much more computationally efficient - using Garry Kasparov vs Deep Blue as an example, Kasparov could not have been searching more than 3-5 positions per second, while Deep Blue, a supercomputer with 480 custom ”chess processors”, searched about 200 million positions per second 1 to play at approximately equal strength (Deep Blue won the 6-game match with 2 wins, 3 draws, and 1 loss).

How can a human searching 3-5 positions per second be as strong as a computer searching 200 million positions per second? And is it possible to build even stronger chess computers than what we have today, by making them more computationally efficient? Those are the questions this project investigates.

[Lai was tapped by DeepMind as a researcher last year]

But what I'm interested in at the moment is the decision speed in these matches:

- What was the average time to make a move in the AlphaZero vs. Stockfish match?


Posted 2018-02-02T22:41:35.850

Reputation: 5 886

@DouglasDaseeco thanks for the vote. (There is a practical intent--have questions about viability of "respectably weak" gofai that can make adequate-but-less-optimal decisions faster than strong ai in certain partisan, economic contexts. Purely academic at this point, but I notice decision speed is not always referenced in research, and for certain algorithmic functions such as high-frequency trading, decision speed is a critical factor. On my own project, it involves acquisition of sectors of memory expressed as cells on a gameboard;) – DukeZhou – 2018-08-07T19:39:43.230



No, ALL computer chess experts were surprised about the outcomes of the match. If you require references, please start a new question.

Your question is simple...


... We evaluated the fully trained instances of AlphaZero against Stockfish, Elmo and the previous version of AlphaGo Zero (trained for 3 days) in chess, shogi and Go respectively, playing 100 game matches at tournament time controls of one minute per move ...

One minute per move. Stockfish would use the one minute if there is more than a single legal move, otherwise it'd move immediately. So the average time for a move for Stockfish is about 57s to 60s.

There is no source code for AlphaZero. However, it's hard to believe the system wouldn't take advantage of the minute it was given. The expected time for AlphaZero should be also 57s to 60s.


Posted 2018-02-02T22:41:35.850

Reputation: 1 375

Thanks for clarifications (re: Chess experts) and the link to the paper. I didn't have time to really research, but the question popped into my head so I figured I'd ask it. As a followup, since your Chess and Chess AI knowledge greatly exceeds mine, can I ask why were the Chess experts surprised? – DukeZhou – 2018-02-05T22:52:37.233

@DukeZhou If you start a new question, we can discuss further. – HelloWorld – 2018-02-05T22:53:27.460