I think additional information that IRL agent needs to recover true reward function is not some prior normative assumptions, it’s non-behavioral data, like “this agent was created by natural selection in particular physical environment, so expected reward scheme should correlate with IGF and imperfect decision algorithm should be efficient in this environment”.
I think additional information that IRL agent needs to recover true reward function is not some prior normative assumptions, it’s non-behavioral data, like “this agent was created by natural selection in particular physical environment, so expected reward scheme should correlate with IGF and imperfect decision algorithm should be efficient in this environment”.