Review for NeurIPS paper: Model Inversion Networks for Model-Based Optimization

Weaknesses: My main concern about the paper is that the proposed method has many moving parts. The method requires the following components: (i) Training a GAN (which is difficult and sensitive to hyperparameters). How many different random seed runs did you try for each expt: what was the variance? What was the sensitivity to hyperparameters? Did you observe issues with mode collapse (or is this not an issue here, because even if we don't generate all possible x, we might still hope to generate some *good* x?) (ii) Approx-Infer: training a model to predict yhat from x, and optimizing (y, z) to maximize the predicted score yhat(xhat(y, z)) while minimizing the disagreement between yhat and y.