NeurIPS 2020

Learning to search efficiently for causally near-optimal treatments


Meta Review

The papers considers the problem of assigning treatments with the side information about the casual graph. The authors propose 3 algorithms and evaluate their performance: an exact algorithm based on dynamic programming, a greedy approximation for high dimensional settings, and a model-free approach motivated by the RL literature. The reviewers for the most part, felt that the application of knowledge of the causal graph to RL is novel and interesting. They were satisfied with the theoretical insights as well as algorithmic contributions and experimental work.