System Level Differentiable Simulation of Radio Access Networks

System Level Differentiable Simulation of Radio Access Networks

Reviewer 1

This paper presents a way of differentiating simulated radio networks so that overall metrics and analyses of the network can be differentiated with respect to parameters behind components of the radio network. My expertise in this area is limited but thtis seems especially useful for parameter optimization in these settings that may otherwise be done with just black box or zeroth-order methods. It would be useful to mention earlier on in the abstract/intro the main insight, intuition, and method of differentiation in this setting since otherwise these details are hidden deeper in the paper.

Reviewer 2

The main contribution of the paper is a differentiable model of cellular network packet loss given assignments of user devices to cell towers. They show that using gradient based optimization with such a model enables them to generate assignments which outperform a simple baseline that assigns each user device to the cell tower with the highest signal strength. The main weakness of the paper is that they don’t compare their technique to any traditional LP/MIP solver based techniques, so it’s difficult to tell if it makes sense to apply back propigation based optimization in their setting or not. I would hope that their technique would at least outperform traditional LP solvers if given a very strict time budget (as one might have in practice in such environments), but it would be good to have results confirming this.