: Reducing emissions, conserving water, and improving energy efficiency to protect the climate and land. Health and Safety
We propose . Unlike fixed hierarchies, AFPM allows the agent to learn a factorization of the policy graph that does not necessarily align with the geometric layout of the environment but rather with the information flow dynamics. This paper applies AFPM to the MRoom domain, demonstrating that arbitrary factorization provides a more robust solution to the "broken bottleneck" problem often seen in stochastic gridworlds. afpm mroom