search

ICSM / ampere

A tool to fit the SED and spectra of dusty objects to constrain, among other things, the dust properties and mineralogy
6 stars 2 forks source link

Search algorithms: Likelihood-free inference with ELFI #42

Closed pscicluna closed 2 years ago

pscicluna commented 3 years ago

As mentioned in issue 14, we need to add more search algorithms, and each one will have its own issue to track progress. This issue will track progress implementing likelihood-free inference (also known as approximate Bayesian computing or ABC) using the Engine for Likelihood-Free inference (ELFI).

LFI is intended for cases where the likelihood function either cannot be computed directly or takes a very long time. We will often fall into the latter category, and LFI will be the method of choice when using a radiative-transfer model or similarly-slow physical model. Although it can accept priors specified more generally, this will be most-easily implemented by using the prior transforms as for nested sampling. I intend to particularly focus on Bayesian Optimisation of LFI, which attempts to find a (non-parametric) function that reproduces the shape of the posterior, so that that can be sampled from.

pscicluna commented 2 years ago

Eventually swtiched to SBI as it provides a wide range of easy to use features. LFI was implemented with Sequential Neural Posterior Estimation in b2300cd0d56d7f8338026aa5503d712c8351761c, making this available. This fulfils the original objectives of this issue, so I will close it and open new issues as required for new features.