DBDaniel Bryarsmsc module / machine-learning

OCOM5200M / Machine Learning

Teaching Models Mars

For the summative project I compared machine-learning approaches for modelling and predicting planetary movement, using NASA JPL Horizons data and the pleasingly awkward geometry of Mars retrograde motion.

I used the project to compare representations as well as models: geocentric and heliocentric data, MLP regression and Gaussian process regression, error measures, hyperparameter ranges and the question of whether a model finds the physics easier when the coordinate system is kinder.

Mars retrograde-style loop plot from machine learning assessment.
Mars is a good reminder that the representation can make the problem look cursed

Assessment

Orbital regression

The project predicted planetary positions from observational data and explored whether heliocentric representation made the learning problem cleaner.

Models

MLP versus GPR

I compared neural-network regression with Gaussian process regression, including kernel choice, periodicity and hyperparameter sensitivity.

Evaluation

Pretty plots were not enough.

The work needed error metrics, validation splits and visual checks because orbit plots can look persuasive while hiding phase or coordinate errors.

Lesson

Inductive bias

A model is not learning in a vacuum. Periodicity, coordinate systems and physical structure all change what “hard” means.

Final orbit prediction plot from the Machine Learning assessment.
Prediction and orbital structure in the final analysis
Initial Gaussian process regression plot.
Gaussian process regression behaving like a useful uncertainty-aware tool
Neural network process diagram from the assessment.
MLP regression pipeline and model structure
Mars loop visualisation.
Representation doing a lot of the intellectual work