Can Bayesian Optimization be the Last Puzzle for Automatic Estimation of Neural Network Discrete Choice Models? An experiment

Abstract

This study investigates the performances of Bayesian optimization (BO) and random grid search methods for tuning neural network hyper-parameters in the context of discrete choice modeling. Specifically, the fully-connected feed-forward (FNN) and alternative-specific-utility neural networks (ASU) are tuned. Results show that BO outperforms random grid search for both FNN and ASU models in terms of out-of-sample log-likelihood. Furthermore, it is illustrated that BO has higher sample efficiency and is relatively more robust to different random initialization. Our experiments show that the Bayesian hyper-parameter tuning framework could accommodate and complement existing neural network models that are cast for automatic utility function specifications, and create a fully automatic estimation workflow.