Comment by ainch
1 day ago
Curious on whether you think this JEPA-style approach could scale to a full game completions?
For reference, standard PPO has been able to beat the game end-to-end with a relatively small network https://drubinstein.github.io/pokerl/
I think it is possible but seems quite difficult to me from this experiment. The hard part is that Pokemon has a lot of different goals, and has many strategies to complete it. You would need a way to encode these intermediary goals in some way temporally that the model needs to learn to complete in order. For example, you need to collect all eight gym badges before you can challenge the Elite Four to win.
You could compose multiple network, one with the goal of learning how to interact with the world through the screenshots. This would probably require a lot of frames of the game from many diverse locations and scenarios--as the network would need to learn how battles works, items (though technically not strictly necessary), and general dialogue and menu interactions.
Then other networks could then try to learn to encode different intermediary goals trained on a bunch of noisy runs that complete that goal. Collecting this data seems tedious and difficult though and is a whole project in itself.
So my intuition is yes, but I don't think simply making my current model bigger would get us there. I see you have experience with world models--what would you try? ;)
That makes sense, thank you! Although it's telling that, with language understanding, a human player is able to discover all the goals for themself, rather than needing to have them specified and shaped up-front. Seems like we still have a long way to go on that front :)
And in terms of what I'd try, I think it's a really hard question haha! I'm interested in the idea of hierarchical world models, where you have something doing prediction on a long-horizon, abstract task level (if I beat this gym I can fight the next one) as well as a short horizon model predicting over individual inputs moment-to-moment. I believe LeCun was actually attached to a hierarchical JEPA paper for robotic control - but it's still very early days.