Parallel Active Inference - PAI

This repositiory contains the supplementary code for the paper "Parallel MCMC Without Embarrassing Failures" available at ArXiv and presented at AISTATS 2022.

Currently, the repostory contain the necessary GP model, mean function, and acquistion function code in the lib directory. The run_toy_4modes.ipynb notebook contains a simple demonstration of the method on the dataset from the paper's section 4.1.

How to install the dependencies

Using conda, the environment.yml file can be used to install all the appropriate libraries. Just run conda env create -f environment.yml and a new ebmcgp can be activated.

Known issues when running on Windows

For Windows users, PyStan might no automatically choose the correct C++ compiler, here's how to fix this:

1. Find the directory of distutils by running:

   import distutils
   print(distutils.__file__)
   # Outputs something like: C:\Users\xxx\miniconda\envs\ebmcgp\lib\distutils\__init__.py

2. Edit or create the file distutils.cfg at the path found above. The new content of this file should be:

   [build]
   compiler=mingw32
   

3. Done!

Reference

de Souza DA, Mesquita D, Kaski S, Acerbi L (2022)
Parallel MCMC Without Embarrassing Failures
To appear in International Conference on Artificial Intelligence and Statistics (AISTATS), 2022. ArXiv:1811.11206

@InProceedings{desouza2022parallel,
  author      = {de Souza, Daniel Augusto and Mesquita, Diego and Kaski, Samuel and Acerbi, Luigi},
  booktitle   = {Proceedings of The 25th International Conference on Artificial Intelligence and Statistics},
  title       = {Parallel {MCMC} Without Embarrassing Failures},
  eprint      = {1811.11206},
  eprintclass = {stat.ML},
  eprinttype  = {arXiv},
  year        = {2022},
}