Review for NeurIPS paper: Learning Linear Programs from Optimal Decisions

Weaknesses: Unfortunately, the authors stop at the rather obvious suggestion that one can apply SQP to the NLP in question. The implementation, while using PyTorch and many rather complicated tools, does not seem to scale beyond very small instances. Perhaps most importantly, the implementation is compared to a rather basic algorithm for generic NLP (COBYLA of Powell), rather than the state of the art in general-purpose DFO or specialised methods for inverse optimisation.