Ask HN: Can we adapt AlphaZero's self-play technique for better human learning?

Since I lack the ML background to debunk this suspicion, I figured I'd let HN debunk it for me. Seeing AlphaZero's success at learning Chess, Shogi, and Go, I was immediately struck with the intuition that the fact that AlphaZero could learn so much from "self-play" should provide some insight into improving human teaching and learning strategies. With the caveat that humans lack AlphaZero's ability to separate themselves into two versions, I can imagine a teaching paradigm that emphasizes simulating competitive activities but playing as both sides. Is something like this at all related to what AlphaZero's doing and are there chess training paradigms that emphasize this type of simulation?

Thanks for being the first to reply. I was worried I'd just get upvotes and no replies!

I think you have two separate points, one with which I agree and one with which I disagree.

First, I agree (and other commentators about AlphaZero seem to as well) that human learning "algorithms" still beat AlphaZero's on per-game ROI.

On the other hand, I disagree that AlphaZero's self-play is no more interesting than a human playing someone better and learning from them. AlphaGo, AlphaZero's predecessor, followed a strategy more like what you described, learning from a large corpus of existing expert chess matches. AlphaZero, on the other hand, requires no training beyond an encoding of the basic rules of chess that it can understand. From there, it bootstraps its understanding of chess without input from experts.

This is the piece I find most interesting, see as potentially useful for the future of human learning, and believe differs from practice with an expert teacher. And so I wonder, can we design learning environments where the learner bootstraps their own understanding from a limited input without continuous feedback from an expert or teacher?

I agree that AlphaZero's per-game learning efficiency is much shorter than a human's (as mentioned in my other reply). The part that interested me more was the fact that it bootstrapped its learning from the basic rules of each game.

Now that I think about it though, one might argue that human learning in a given discipline starts as isolated with feedback only coming from the outside world. This is what we typically call research. But the magic of our education system, when it works, is that we compress the output of this slow process into a faster one and feed it to learners, allowing them to build understanding of knowledge which originally took generations to discover. Riffing off Matt Might's illustrated depiction of a PhD (http://matt.might.net/articles/phd-school-in-pictures/), expanding the circle of knowledge is exponentially slower than getting close to the edge.

> And so I wonder, can we design learning environments where the learner bootstraps their own understanding from a limited input without continuous feedback from an expert or teacher?

Yes, and we do it all the time.

We can just do a lot better with continuous feedback. (And AI probably could, too, if experts that could communicate fast enough not to be a huge drag on the AIs training cycles were available. But since with current technology once we've trained an AI of the type we can make today, we can replicate it, that's not really important; if ever developed AIs that depend on reconfigurable hardware without trivially extractable state, that may change.)

AlphaZero plays games with (1) perfect information and (2) well-defined winning conditions. Neither of these hold for most human-learning scenarios.

I can imagine that a healthy dose of probability theory (and probably more advanced stuff I don't know about[1]) might improve (1), but (2) is going to keep computer scientists and philosophers and ethicists arguing for quite a long time. :)

[1] get the joke, eh? eh? eh?