Chainsail resources and documentation

Examples, documentation and other additional resources related to the Chainsail sampling web service (https://chainsail.io).

Chainsail: a web service to sample multimodal probability distributions

Chainsail is a web service which helps you sample from multimodal probability distributions. In the context of Bayesian statistics, they arise in the case of unidentifiable parameters which are due to some symmetry in the model or if you have ambiguous data.
While this early version of Chainsail is targeted towards experienced Markov Chain Monte Carlo (MCMC) practitioners, it is still designed to be user-friendly and in this repository, we provide some documentation on how to use Chainsail and the algorithms at work behind the scene.

Usage

Learn how to use Chainsail by checking out the walkthrough here. Most importantly, you will want to learn how to define your own probability distribution.

The algorithms behind Chainsail

Chainsail implements several important algorithms, which we describe here in not too much detail:

• Replica Exchange: Chainsail's main ingredient that allows you to sample multimodal probability distributions
• Automatic Replica Exchange tuning: Replica Exchange requires setting a kind of "temperature" schedule, which Chainsail automatically determines for you
• Hamiltonian Monte Carlo: While Replica Exchange takes care of global sampling, meaning it helps to discover all modes of your probability distribution, local sampling algorithms like Hamiltonian Monte Carlo sample well within a single mode. Chainsail currently only implements a very simple form of Hamiltonian Monte Carlo.

The chainsail_helpers package

This repository also contains the source code for the chainsail_helpers package. It defines the interface for Chainsail-compatible probability distributions, PPL-specific implementations of them and provides helper scripts.

Questions?

Shoot us an email!