Inference using Gaussian processes in animal movement modelling

Paun, Ionut Alexandru (2022) Inference using Gaussian processes in animal movement modelling. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of 2022PaunIonutPhD.pdf] PDF
Download (17MB)

Abstract

In recent years, the field of movement ecology has been changed dramatically by the capacity to collect accurate high-frequency telemetry data. In this thesis I present new statistical methods scalable to very large volumes of data being generated as there is a problem of scale dependence in most popular animal movement models. Popular and widely used movement models in ecology are discrete-time movement models, where animals’ positions are observed at discrete times. However, discrete-time models do not perform well when problems such as missing or irregular data are present. A remedy to the inefficiency of discrete-time movement models is to use continuous-time movement models, however the formulation of continuous-time movement models is often difficult and hard to interpret. In this thesis, I first focus on discrete-time movement models, where through a study I illustrate one of the problems that discrete-time movement models pose - the specification in advance of the discretisation time-step. I then move on to probabilistic methods, widely used in the machine learning community, Gaussian processes (GPs), and I show that they are equivalent to many continuous-time movement models. Given that the primary goal of machine learning methods is to learn from large scale datasets, using robust continuous-time movement models such as Gaussian processes is highly advantageous for multiple reasons. These include their flexibility in choosing various covariance functions, their scalability to large datasets and their ability to analyse data, infer parameters of interest and quantify uncertainty within a nonparametric Bayesian approach. I extend the standard Gaussian process (GP) into a non-stationary hierarchical Gaussian process, where both the movement process and the dynamic parameters of the movement model are Gaussian processes, which allows for increased flexibility to a wide range of behaviour modes that animals can exhibit. Throughout this thesis, I implement Gaussian processes on simulated and real tracking data using statistical libraries such as TensorFlow, which provide an accessible way to implement the model and gain access to GPU/HPC-accelerated machine learning libraries. I perform inference using optimisation methods such as maximum-a-posteriori (MAP) estimation, approximate sampling based inference methods such as Markov Chain Monte Carlo (MCMC) and variational inference methods on both synthetic and real datasets.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Colleges/Schools: College of Science and Engineering > School of Engineering
Supervisor's Name: Torney, Dr. Colin and Husmeier, Prof. Dirk
Date of Award: 2022
Depositing User: Theses Team
Unique ID: glathesis:2022-83128
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 14 Sep 2022 15:28
Last Modified: 14 Sep 2022 15:28
Thesis DOI: 10.5525/gla.thesis.83128
URI: https://theses.gla.ac.uk/id/eprint/83128

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year