3 ms·
Interesting. I'm not a deep learning guy, but from what I can gather, the new auxiliary tasks are to be rewarded for "pixel changes" and "network features". I
by TFortunato 10y ago
Interesting. I'm not a deep learning guy, but from what I can gather, the new auxiliary tasks are to be rewarded for "pixel changes" and "network features".
I haven't nearly finished reading the paper, but is it safe to say this is similar (at a very high level), to a type of "novelty search", where the agent is searching not only for a policy that is directly accomplishing the task at hand, but also for novel stimulus (in the case of pixel changes), and novel internal states (features, or maximally activated hidden nodes in the language of the paper), and that the benefit of this would be to more easily find relevant features that could be useful in the "big picture" task, and maybe not get as stuck in a non-optimal policy?
(I may be understanding this completely wrong...just an embedded guy looking to get more into this world, haha)
- TFortunato 10y agoQuick reply to my own comment. But another reason it seems this is helpful is that in many reinforcement learning tasks, like games, rewards are few and far between, so these goals also give the agent oncentive to keep trying new things / learning about the environment / task in the presence of sparse rewards?
- mmusson 10y agoIt seems to me like an attempt to model curiosity.
- TFortunato 10y agoExactly what I was thinking, and put much more succinctly!