Adaptive Sampling For Efficient Online Modelling

Abstract

This thesis examines methods enabling autonomous systems to make active sampling and planning decisions in real time. Gaussian Process (GP) regression is chosen as a framework for its non-parametric approach allowing flexibility in unknown environments. The first part of the thesis focuses on depth constrained full coverage bathymetric surveys in unknown environments. Algorithms are developed to find and follow a depth contour, modelled with a GP, and produce a depth constrained boundary. An extension to the Boustrophedon Cellular Decomposition, Discrete Monotone Polygonal Partitioning is developed allowing efficient planning for coverage within this boundary. Efficient computational methods such as incremental Cholesky updates are implemented to allow online Hyper Parameter optimisation and fitting of the GP's. This is demonstrated in simulation and the field on a platform built for the purpose. The second part of this thesis focuses on modelling the surface salinity profiles of estuarine tidal fronts. The standard GP model assumes evenly distributed noise, which does not always hold. This can be handled with Heteroscedastic noise. An efficient new method, Parametric Heteroscedastic Gaussian Process regression, is proposed. This is applied to active sample selection on stationary fronts and adaptive planning on moving fronts where a number of information theoretic methods are compared. The use of a mean function is shown to increase the accuracy of predictions whilst reducing optimisation time. These algorithms are validated in simulation. Algorithmic development is focused on efficient methods allowing deployment on platforms with constrained computational resources. Whilst the application of this thesis is Autonomous Surface Vessels, it is hoped the issues discussed and solutions provided have relevance to other applications in robotics and wider fields such as spatial statistics and machine learning in general.