Date of Award

Spring 1-1-2015

Document Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

Eric W. Frew

Second Advisor

Nisar Ahmed

Third Advisor

Dirk Grunwald

Fourth Advisor

Douglas Sicker

Fifth Advisor

Michael Mozer


This dissertation develops a nonparametric, computationally efficient method for modeling the airborne communication environment for small unmanned aircraft systems (sUAS). Transfer learning for Gaussian process regression (GPR) allows the communication model to adapt to the spatial and temporal variations within an environment, and to the variations that arise across UAS hardware and missions.

Environment-specific radio frequency (RF) variations are learned by augmenting a parametric path loss model with a nonparametric Gaussian process (GP), which captures geospatial and time-varying characteristics of signal strength measurements. This dissertation assesses the performance of GP-based communication models through cross validation on 50 sets of real light measurements collected using two different frequencies, three different airframes, and employing static and mobile transmitters. Measuring the performance using root mean squared error (RMSE) as well as mean standardized log loss (MSLL) evaluates both the predicted estimate and its uncertainty, and shows that the GP models improve prediction accuracy over the path loss model, with the spatio-temporal GPs improving over the spatial GPs.

The value of GP-based communication modeling is further demonstrated through integration with UAS data ferrying to opportunistically learn geospatial variations in RF measurements and use them in communication link scheduling. The iterative ferry-and-learn system is analyzed through a simulation study, showing ferry achieves 80% of optimal within 4 iterations, and 93% after 9 iterations, as the GP is able to converge quickly to the true radio frequency environment. Comparison with parametric least-squares learners in two extremes of RF scenarios shows that the GP better captures the stochasticity of the environment, especially in complicated cases. This demonstrates that learning RF variations using GPs provides a significant boost to the performance of communication-aware UAS applications.

Because all communication in the environment is affected by the same factors, previously learned GPs contain information that is relevant while learning communication models for subsequent missions and on different UAS platforms. This dissertation proposes forward adaptive transfer learning for Gaussian process regression, FAT-GP, which allows previously learned GP models to be adapted forward as potential sources of knowledge for future learning tasks, which can be especially valuable when limited training data is available for the new task. FAT-GP combines the source task's previously learned model, the source task's training data, and target task's training data to learn the target hyperparameters as well as the correlation between the two tasks. This extension to GPR not only generalizes transfers between GPs using different kernels, but also results in amortization of the training cost.

Such an adapt-and-update framework is in keeping with the philosophy of lifelong learning, and is valuable in UAS and other robotics missions, especially when operating in unstructured and unexplored environments.