• About
  • Organizers
  • Location
  • Code of Conduct
  • Team
  • Gallery
  • Sponsors
Categories
  • Statistics

Project 4:

cosmoabc: Likelihood-free inference via Population Monte Carlo Approximate Bayesian Computation

Due to the increasing amount of data and their complex modelling in all areas of astronomy and cosmology, more and more computational power is required in order to explore larger parameter spaces whose internal correlations can often be impractically complicated or unknown. Thus, obtaining a statistical tool which bypasses the need of fully evaluating the likelihood is imperative.

This project presents cosmoabc, a Python implementation of Population Monte Carlo Approximate Bayesian Computation (PMC-ABC) algorithm with adaptive importance sampling. cosmoabc is the first such implementation targeted to the astronomy and cosmology community. The cosmological simulations are done through a connection with the Numerical Cosmology library (numcosmo), but the code is flexible enough for user-specified distance, simulation and prior functions.

We demonstrated the power of cosmoabc in estimating posterior probability distributions in two situations: a simple toy model and a complex cosmological simulation of a Sunyaev-Zeldovich survey. In both cases, we show how cosmoabc allows a good approximation of the true posterior probability distribution with a fairly simple and user friendly interface.

Full citation: Ishida et al., 2015, Astronomy and Computing, Volume 13, p. 1-11

This project is a result from COIN Residence Program #1 – Lisbon, Portugal/2014.

Pre-print – Bibtex – Package

Team:
Emille E. O. Ishida, CNRS/UCA (France)
Sandro Vitenti, Louvain U. (Belgium)
Mariana Penna-Lima, U. Brasilia (Brazil) and LAPP (France)
Jessi Cisewski, Yale U. (USA)
Rafael S. de Souza, U. North Carolina (USA)
Arlindo M. M. Trindade, U. Porto (Portugal)
Ewan Cameron, Oxford U. (UK)
Vinicius C. Busti, U. Sao Paulo (Brazil)

Behaviour of a sufficient summary statistics as a function of the parameters of our toy model.

fb-share-icon
Tweet

Related posts

The Gaussian components projected onto the BPT diagram. For each component the thick lines represent 68% and 95% confidence levels, respectively.

23/04/2018

9 – Gaussian Mixture Models


Read more
23/04/2018

1 – Binomial regression


Read more

The 2D probability density of the predicted redshift from the GLM fit vs. the spectroscopic redshift (central plots) of DSS sample. The upper and right subplots depict the redshift distribution along photo-z and spectroscopic redshift, respectively.

23/04/2018

2 – Gamma Regression


Read more
© 2018 COIN