MCMC for posterior distribution sampling

Satwik Kottur and Krishna Pillutla, Carnegie Mellon University

This project is a part of 10-708: Probabilistic Graphical Models, Fall 2015, in requirement for the course completion.

The idea is to handle non-smooth energy functions in the setting of Hamiltonian dynamics for Monte Carlo Markov chain (MCMCM) sampling. Hamiltonian Monte Carlo (HMC) methods evolve a set of differential equations and non-smooth energies do not fit in, as in. The report provides further details on the various strategies adopted to solve the problem.

Parts of the code are adapted from two sources:

1. Hamiltonian Monte Carlo MATLAB Implementation
2. Tianqi Chen's Implementation for Stochastic Gradient-HMC