NeurIPS 2020

Influence-Augmented Online Planning for Complex Environments

Meta Review

The paper addresses a problem of computationally expensive forward simulation in POMDPs. The authors propose a POMDP solver that uses influence-based attention to cheaply estimate substates, reducing overall cost of the forward simulation. There is very little work in reducing the cost of forward simulation, which makes this work interesting and important. The method is well founded, with somewhat weak in the evaluations. The evaluations shows significant benefit in two domains, but omit in-depth analysis of the algorithm. In the light of the after the authors' responses and reviewers' discussion, the authors should include: - The discussion from their response on the differences to the exogenous and endogenous factors in POMDPs, and clearly explain that the influence-based attention is more general.