Interpretable Methods for Adaptive Route Improvement Models Based on Behavioral Trajectory Prediction

Abstract

The emergence of Artifical Intelligence (AI) and Machine Learning (ML) have transformed various sectors by enabling machines to learn from data, recongnize patters, make decisions, and perform taskts requiring human intelligence. In the automotive domain, Smart EV Routing by WirelessCar optimizes electric vehicle routes using real-time and user data. One potential modification to such a system would be the integration of a model that autonomously determines the route based on historical data and environmental conditions. With the use of interpretability, one can ensure that such a system works and gives accurate predictions reflecting the user driven data. This thesis develops a framework to integrate such a system while focusing on interpretable trajectory prediction models and interpretability techniques. It also tackles the complexity of multivariate data and addresses issues related to data scarcity. While the focus has primarily been on models learned in simulation, the framework is designed with future applications in mind, aiming to support the development of methods with similar constraints and requirements. The results show that the existing interpretability methods are inadequate for scenarios involving time series data where multiple variables (multivariate settings) affect the outcomes over time. Attempting to aggregate these methods in multivariate settings results in loss of information and increased complexity. This renders these methods impractical for dynamic domains such as car trajectory prediction, this was also evident in the result. To address these issues we developed Interpret Multivariate Timeseries (IMT), a method developed specifically for multivariate settings, as a true black box explainer. The results also show that our framework integrates multivariate forecasting models, interpretability methods, adaptability, and data generation, enabling interpretable and adaptive route improvements based on behavioral trajectory prediction.