Evolutionary Optimization for Neural Network-Based Discrete Choice Modeling: Enhancing Stability and Efficiency

The integration of theory-driven discrete choice modeling (DCM) with neural networks has demonstrated promising advances in choice behavior analysis, yet existing hybrid approaches face persistent challenges in parameter stability, overfitting, and computational efficiency. This study presents ResLogit Plus, a novel framework that enhances the synergy between residual neural networks and discrete choice models through genetic algorithm (GA) optimization. By leveraging GA for parameter initialization and hyperparameter tuning, our approach significantly improves model stability and computational performance while maintaining interpretability. We introduce a comprehensive validation framework incorporating regularization techniques and bootstrap-based validation to ensure robust parameter estimation and mitigate overfitting risks. Empirical validation is conducted using three datasets conducted in different countries: the Swiss Metro dataset (Switzerland), the Carpooling dataset (Switzerland and Germany), and the Greenhouse Gas Emissions dataset (Canada). The results reveal that ResLogit Plus achieves superior predictive accuracy compared to both the original ResLogit and traditional Multinomial Logit (MNL) models, while demonstrating enhanced parameter stability and reduced computational overhead. The framework effectively addresses key methodological challenges, including correlated alternatives and utility estimation noise, thereby advancing the field of discrete choice analysis through a balanced integration of predictive power and theoretical rigor.