Learning Salon: Katja Hofmann, Deep RL for Games

February 27, 2021

Very quick notes from Katja’s Learning Salon talk on Deep RL for Games at MS Cambridge.

Discussion

Are games “just something that conforms to RL algorithms?” No. But in games, human players and AI are at similar footing.

What is an “RL agent”? Whatever we can model. [Yes, pretty sure that’s exactly what Sutton&Barto mention in their book.] John: we can’t ascribe “agency” to the weather. [Can’t we?]

Why games? Combination of motivation, control and flexibility. Easy to collect data, can be setup in a multitude of ways. Virtual environments can be made “safe” [assuming transfer to real world works].

Games where you can’t define a goal beforehand? Curiosity-driven exploration, automated curriculum learning, etc. Katja is hopeful on meta-learning.

Katja: it makes sense to assume you can define anything as a “reward”, but the rewards themselves are varied (e.g. humans can learn based off of “well done”).

[NOTE: a lot of the things they are discussing seem to be rehashed from Minsky’s Society of Mind.]

Learning to be a good opponent seems to be easier than learning to be a good teammate. [Makes sense, it’s easier to define the inverse optimization function.] Maybe make mistakes on purpose and then force other agent to compensate?

How about free play mode? Katja didn’t make experiments there yet. Some ideas on how to force agents to find their own goals: social environment. John: “Everything is a version of something before it” [Well yes, the world is compositional, only physics is true.]

Darwin: don’t conflate story about how you got to X vs. how X works. E.g. there is a story on how you evolved a brain, not how it works.

Jovo: There is no objective in evolution. [Not sure about that. Isn’t human learning just optimizing for survival given some “natural” inductive biases?]

John: Our learning not completely different from animals, more like “average animal + something else”. Brad: boredom as creativity signal [Again something Minsky discusses].

Zenna asked something like “is creativity an emergent property of RL-model-based optimization” (?) [I think so, but eehhhh.]

John: Games are a bit more like an engineering problem, vs. philosophical questions. “Why is there a river here?” vs. “How do I build a bridge over this river?” [I think John is actually hinting at intrinsic motivation and meta-learning. We are actually trying to model these questions with RL.] There’s no reason to be optimistic about it without any concrete proof, and it seems that we don’t have proof right now.

David: The agents are quite passive.

Have you modeled something like joy in the agents? Nope; what does joy mean for a human player?

Penguins are fun

Learning Salon: Katja Hofmann, Deep RL for Games - February 27, 2021 - tkukurin