5
$\begingroup$

I don't know much about AI and am just curious.

From what I read, AlphaZero/MuZero outperform any human chess player after a few hours of training. I have no idea how many chess games a very talented human chess player on average has played before he/she reaches the grandmaster level, but I would imagine it is a number that can roughly be estimated. Of course, playing entire games is not the only training for human chess players.

Nonetheless, how does this compare to AI? How many games do AI engines play before reaching the grandmaster level? Do (gifted) humans or AI learn chess faster?

$\endgroup$
0
1
$\begingroup$

According to Table S3 of the AlphaZero paper (p. 15)

enter image description here

AlphaZero was trained for 9 hours and, during these 9 hours, it played 44 million games of chess.

According to this Wikipedia article, the longest human lifespan is that of Jeanne Calment, who lived to age 122 years and 164 days.

Let's assume that humans cannot live more than 123 years (which is a reasonable assumption, although this record could eventually be broken). Let's also assume that a chess game lasts at least 10 minutes, which means that you can play at most 6 games in 1 hour, which means that you can play at most $6*24 = 144$ games in one day (assuming that you never sleep, which is, of course, impractical, but I'm just trying to show you an upper bound). Let's say that a year has 365 days. So, here is roughly the maximum number of games that a human could play $$ 6*24*365*123 = 6464880 \tag{1}\label{1}, $$ which is smaller than 44 million games by a factor of more than 6, i.e. any human could at most play 1/6 of the games that AlphaZero played, and \ref{1} is a very loose upper bound that doesn't take into account that humans need to sleep, eat, and do many other things.

So, humans learn to play chess a lot slower than AlphaZero. This is not surprising at all, given that computers can perform calculations a lot faster than us (that's why they are called computers), and this has been the case for many years. We (humans) just made computers make the right calculations for them to approximately play chess better than us. That's it.

$\endgroup$
2
  • 1
    $\begingroup$ Thank you for the answer! So we learn a lot slower in absolute time, but the best chess players do very well in comparison in number of games played. It's frightening to think that one day AI may learn faster even in number of games played. $\endgroup$
    – 220284
    Feb 5 at 7:46
  • $\begingroup$ @220284 Good observation. We'll probably need to develop some new insights in transfer learning for that to happen, but I agree with you that, when AIs can learn from fewer examples/experience than us, that will definitely be a good sign of "intelligence" (or, more precisely, the ability to "generalize" or "transfer knowledge" that we often associate with intelligent systems). $\endgroup$
    – nbro
    Feb 5 at 14:02
-1
$\begingroup$

A simple Google search such as https://www.quora.com/How-many-games-did-Alpha-Zero-played-against-itself-during-its-four-hours-training will tell you about AlphaZero took about 44 million games to achieve in just hours.

No human will of course can do it.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.