Yeah, but the method of learning is interesting - unsupervised training in a VR world. As CPUs become faster, than these type of problems will fall much faster.
I am not sure it is completely unsupervised, iirc they said it was "coached". I don't think a bot can learn by itself to fake razes/perfect creep blocking by itself.
They talked about this twice, the exact text of both follows
> We didn't program it to understand the rules of Dota, we just let it play lifetimes of 1v1 against itself and coach it on what we thought was good or bad.
> We've coached it to learn just from playing against itself. We didn't hard code in any strategy. We didn't have it learn from human experts. Just from the very beginning it just playing against a copy of itself. It starts from complete randomness and then it makes very small improvements and eventually reaches the pro level. [... goes on to talk about how it's first "improvement" was to just never leave the base because it kept dying]
I think you're reading too much into the word coached. Particularly the second quote implies that it wasn't supervised at all.
My personal explanation for "Fake razes" is that it isn't a fake out to the bot, it's a technique to lower the casting time to zone out the opponent/hit faster if the opponent chooses to dives. Seems like a pretty easy trick to learn.
Creep blocking is just learning "positioning myself here (in front of the creeps) tends to improve my chances of winning", where "here" isn't far from where it would be positioning itself without creep blocking. It's a minor adjustment to the behavior with a substantial effect on the game outcome, of course it learned it.
In this post, a member of the OpenAI team characterizes this form of coached self-play as being supervised. If you are assigning a value to an action that is not an explicit goal of the game and is not a consequence of violating the game's rules, it would seem to fit the definition of 'supervised'.
