To the left, you see a trained agent playing a level of a game. To the right, you see the same playthrough from an agent-centric perspective: cropped, translated, and rotated with the agent in the center. Which perspective is the best input for the agent? https://arxiv.org/abs/2001.09908 pic.twitter.com/7bCtBp8xUG
-
-
Also, can it? It is hard to imagine that a neural network of just a few layers could actually implement the transformations necessary to even understand where things are relative to the agent, so that the policy can be location-independent?
Prikaži ovu nit -
It is possible that the standard paradigm of a neural network with a handful of layers learning to master, say, Atari games from a static third-person view is actually impossible. That is, it doesn't learn any general playing skills. It learns some kind of stimulus-response table
Prikaži ovu nit -
In any case, even if this is possible in principle, it seems that the way we represent the input makes a lot of difference for the generality of skills that can be learned in practice.
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
-
-
Yeah this strikes me as very much a case of "Why would you *want* it to be general, though" from a developer's perspective. Like yeah, I don't need my AI to learn who they are, it makes sense that I'd teach them that!
-
But I guess the aim of a lot of RL work on games isn't really to make games better (present company excepted of course).
- Još 3 druga odgovora
Novi razgovor -
-
-
If there are distractor concepts we don't want the agent to learn (e.g. which map; what absolute agent position) I'd think to use adversarial reprentation learning. Add a penalty for when the distractor concepts are predictable from the learned repr. Train like in GANs.
-
I think a lot of researchers are looking into learning representations from the environment. But this requires multiple level examples if you want a general representation, or labeled data in the distractor example. I do think this work supports moving to learned representations.
- Još 4 druga odgovora
Novi razgovor -
-
-
the partially observed aspect seems relevant too; if agent is handed a 'god's eye view' there is less incentive to explore and develop a more organic sense of geography
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.