Nested Sampling for Javascript applications

I demonstrate how to build useful, interactive statistics tools using model comparison between low-dimensional models and parameter estimation.

Features of the framework

• Web tools are interactive, the user can choose or enter data
• Web tools are easily accessible, the user does not have to compile or download anything (low setup cost) and can play with data without effort.
• With RadFriends / Nested Sampling (http://arxiv.org/abs/1407.5459) you do not need to worry about convergence, this is handled safely.
• You can also do parameter estimation -- posterior samples are computed for you.
• Beware that RadFriends becomes inefficient for high-dimensional problems. Depending on your specific problem, that may become limiting already at 5 dimensions. This implementation is aimed for simple, low-dimensional problems (line fitting, comparing distributions)

Demos

Try it out: https://johannesbuchner.github.io/ultranest-js/twodist.html

I show an application where the user can enter two data sets. Two models are compared. In the first model, all data are generated from a normal distribution with unknown parameters (2 parameters). In the second model, the two data sets are each generated from a normal distribution with unknown parameters (4 parameters). After clicking the start button, the marginal likelihood (evidence, Z) of both models is computed, and compared. Please take a look at the source files.

Try it out: https://johannesbuchner.github.io/ultranest-js/gaussian.html

In this application, the user can enter measurement values with uncertainties. A gaussian is fitted to the data, taking into account the limited number of measurements and the measurement uncertainties.

Please fork and build your own application!

Files

• twodist.html: HTML interface, with plotting
• js/worker.js: WebWorker interface
• js/problem2.js: Implementation of the likelihood and prior function
• js/plot.js: Some plotting
• js/js_nested2.js: Generic Nested Sampling implementation using RadFriends
• gaussian.html: HTML interface for gaussian demo, with plotting
• js/gaussworker.js: WebWorker interface for gaussian demo

Author

Some files are based on the BEST tool (http://www.sumsar.net/best_online/), and are attributed appropriately. Otherwise, the code is written by Johannes Buchner (C) 2014-2018, and licensed under AGPLv3 (see LICENSE file). If you require a different license, please contact me.