Machine Learning Cosmic Expansion History

TL;DR

This paper applies machine learning techniques to cosmic chronometer data to estimate the Hubble constant and constrain cosmological models, demonstrating the effectiveness of Bayesian ridge regression and model compression.

Contribution

First application of machine learning to cosmic expansion history using 30 chronometers, identifying Bayesian ridge regression as optimal for this task.

Findings

Bayesian ridge regression outperforms other algorithms in this context.

Estimated Hubble constant as 65.95 km/s/Mpc with uncertainties.

Many cosmological models with more than 3 parameters can be ruled out or compressed.

Abstract

We use the machine learning techniques, for the first time, to study the background evolution of the universe in light of 30 cosmic chronometers. From 7 machine learning algorithms, using the principle of mean squared error minimization on testing set, we find that Bayesian ridge regression is the optimal method to extract the information from cosmic chronometers. By use of a power-law polynomial expansion, we obtain the first Hubble constant estimation $H_0=65.95^{+6.98}_{-6.36}$ km s$^{-1}$ Mpc$^{-1}$ from machine learning. From the view of machine learning, we may rule out a large number of cosmological models, the number of physical parameters of which containing $H_0$ is larger than 3. Very importantly and interestingly, we find that the parameter spaces of 3 specific cosmological models can all be clearly compressed by considering both their explanation and generalization abilities.