Machine Learning, 2nd Edition

Chapter 18

Gaussian Processes

The supervised machine learning algorithms that we have seen have generally tried to fit a parametrised function to a set of training data in order to minimise an error function. This function is then used to generalise to previously unseen data. Some of the differences between the methods have been the set of model functions that the algorithm can use to represent the data; for example, the linear models of Chapter 3 and the piecewise constant splines of Chapter 5. However, if we do not know anything about the underlying process that generated the data, then choosing an appropriate model is often a trial-and-error process.

As a very simple example, Figure 18.1 shows a few datapoints. If we assumed that these were ...