Capturing Human Route Preferences from Track Information: New Results

Abstract

In previous work, we described G2I2, a system that adjusts the cost function used by an off-road route planning system in order to more closely mimic the route choices made by humans. In this paper, we report on an extension to G2I2, called GUIDE, which adds significant new capabilities. GUIDE has the ability to induce a cost function starting with a set of historical tracks used as training input, with no requirement that these tracks be even close to cost-optimal. Given a cost function, either induced as above or provided from elsewhere, GUIDE can then compare planned routes with the actual tracks executed to adjust that cost function as either the environment or human preferences change over time. The features used by GUIDE in both the initial induction of the cost function and subsequent tuning include time-varying meta-data such as the temperature and precipitation at the time a given track was executed. We present results showing that, even when presented with tracks that are very far from cost-optimal, GUIDE can learn a set of preferences that closely mimics terrain choices made by humans.