Review for NeurIPS paper: Influence-Augmented Online Planning for Complex Environments

Weaknesses: The major concern is that the idea of exploiting "influences" of domain variables to reduce the state space of POMDPs is not new. In the literature, those variables that only indirectly influence agent behaviors are referred to as exogenous variables. The following are two papers that studied this idea. The RNN-based influence learning is new within the literature, while the following two papers have studied other reasoning and learning methods to incorporate exogenous variables into POMDP-based action selection processes. Zhang S, Khandelwal P, Stone P. Dynamically constructed (PO) MDPs for adaptive robot planning.