Learning Routing Preferences for a Personalized Navigation System

Abstract

Current-day navigation systems do not offer highly-adaptive personalized route recommendations to users, which poses a problem to many such as new drivers or during times of extreme weather. In this thesis, we present two versions of an app designed to provide an end-to-end framework for personalized navigation that adapts to user preferences in real-time. Our second app version proposes a novel approach to personalized route generation by applying state-of-the-art techniques from Reinforcement Learning from Human Feedback (RLHF) to approximate cost functions aligning to human preferences. We begin by introducing the concepts and techniques used in our approach to personalized route recommendations, before presenting our framework. We find through a preliminary user study with Princeton University students (N=7) that participants preferred the preference-based learning version of our app and they report higher values of ease of use and app capability for this version.