RIPPIL

ABSTRACT

RIPPIL investigates the applicability of interactive learning to enhance automated route planning. We developed a capability that learns and exploits a user’s latent reward function through pairwise comparison feedback on presented aerial routes created using a mature, fielded automated route planner. Initial results with synthetic reward functions show that the presented routes improve over time (i.e., more closely align with the feedback) and the generation of route pairs using active strategies speeds up the interactive learning process.

WHY?

Many automated mission planning tools have been developed and matured over the past couple of decades, but often do not capture everything a user desires, leading to non-use or time-consuming manual alteration of plans. By combining the computational strengths of these automated planners with the expert user’s intuition and domain expertise, we strive to achieve higher quality plans in a shorter time.

WHAT?

RIPPIL applies preference learning to automated low-observable aerial route planning by iteratively modifying route planner inputs and collecting user comparison feedback on presented route pairs. While pairwise preference learning has been implemented on low-dimensional problems in the literature, it has not been extensively evaluated for high-dimensional input spaces and on real user preference data. We will collect and analyze data from user experiments to determine the benefits of interactive learning for aerial route planning.

HOW?

RIPPIL’s Gaussian Process-based preference learning prototype has been tested using synthetic reward functions to simulate preference decisions, and will be tested with real user preferences.