Post-hoc explanation methods for deep neural networks in choice analysis

Abstract

Large-scale traffic simulation models are a crucial tool for simulating and evaluating different transport solutions. However, due to the scale and complexity of these models, numerous parameters exist that can significantly influence their outputs. The problem of estimating these parameters is referred to as the Dynamic Traffic Assignment (DTA) calibration problem. After more than 30 years of research, several algorithms have been proposed that can - with a certain degree of success - address this challenge, even for large instances or in the presence of noisy data. Two challenges, however, remain critical today and are addressed in this paper. From a purely methodological perspective, DTA calibration is a highly under-determined problem, meaning multiple plausible solutions exist. This is particularly relevant when calibrating demand parameters. Therefore, in this paper, we propose two techniques inspired by the field of computer science that allow for enhancing robustness: bagging and Stochastic Parameter Averaging (or SPA). The second contribution of this research is more practical. While many algorithms have been proposed, the source codes of these algorithms are often not shared with the scientific community. As a consequence, most papers still use as a benchmark model the SPSA, an algorithm proposed roughly 30 years ago. Therefore, this study introduces an end-to-end open-source framework for DTA calibration. The model can calibrate supply and demand parameters, include state-of-the-art optimizers (W-SPSA, SPSA, Bayesian Optimization), an auto-tuning option to calibrate their parameters, and the bagging/SPA extension already mentioned. The conceptual framework proposed in this research is general and includes a few algorithms already. It is currently linked with the open traffic simulator SUMO to demonstrate its effectiveness. Researchers can use this framework as a benchmark or extend it using new simulators and optimizers. The method is tested both in controlled settings, as well as using the real-world large scale network of Munich.