Applicability to non-RL problems?

[shouldsee]

Thank you for this exciting work!!! Just wondering have you tried this methodology in non-reinfrocement settings? Would you anticipate a diffcult adaptation of this methodology?

[lerrytang]

Hey, thanks for the question.
I've mostly tried the method in RL settings (multi-agents, collective intelligence, etc), but similar methods can be applied to set classification (e.g., SetTransformer).
Do you mind sharing the target settings you are considering? That may allow me to give some better suggestions :)

[shouldsee]

Hey, thanks for the question. I've mostly tried the method in RL settings (multi-agents, collective intelligence, etc), but similar methods can be applied to set classification (e.g., SetTransformer). Do you mind sharing the target settings you are considering? That may allow me to give some better suggestions :)

Thanks for the reply. Unfortunately I don't have a specific problem at the moment, but was thinking about data mode inference that process images and audio using different set of weights. But the setting would be a bit different, since changing the weight might not be preferential, because this means the model could not process audio and image simultaneously

It's quite curious that NN could adapt to a permutation. Is there evidence that a fully trained network would adapt to a permutation more quickly than a random network?