Given a dataset of 1D continuous-valued input-output pairs cw1a.mat, Gaussian Process models with different covariance functions were individually fit to the data and their performance compared with respect to marginal likelihood. For each model, its hyper-parameters had been chosen to maximise the marginal likelihood, so as to avoid overfitting.
Below shows the data fits for isotopic squared exponential and periodic covariance functions respectively:
Gaussian Process had also been used to perform regression on a 2D-input 1D-output dataset cw1e.mat. The best result is demonstrated below, where the red surface maps the observed data; black surface gives the prediction:
Demo above is part of my university coursework. Details of problem statements and analysis can be found in Problems.pdf and Report.pdf respectively.