POTMMCP

This repository contains the implementation of the Partially Observable Type-based Meta Monte-Carlo Planning (POTMMCP) algorithm from the paper: Combining a Meta-Policy and Monte-Carlo Planning for Scalable Type-Based Reasoning in Partially Observable Environments.

Installation

The library is implemented in python 3.8 and has the following main dependencies:

1. Pytorch >=1.11
2. Rllib == 1.12
3. posggym (comes with repo)
4. posggym-agents (comes with repo)

As with any python package we recommend using a virtual environment (e.g. Conda).

Installation of potmmcp requires cloning the repo then installing using pip:

git clone git@github.com:Jjschwartz/potmmcp.git

# install posggym
cd potmmcp/posggym
pip install -e .

# install posggym-agents
cd ../posggym-agents
pip install -e .

# install potmmcp
cd ..
pip install -e .

This will install the potmmcp package along with the necessary dependencies.

The codebase

There are two main parts to the codebase:

1. The potmmcp directory containing the potmmcp python package
2. The experiments directory containing scripts and Jupyter notebooks for running and analysing the experiments used in the paper. The results are also store here.

potmmcp

The potmmcp python package contains a few main parts:

1. baselines - implementation code for the different baselines used in the paper
2. plot - code used for generating plots and running analysis
3. run - code for running and tracking experiments
4. tree - the implementation of the POTMMCP algorithm
5. meta-policy.py - classes and functions implementing the meta-policy
6. policy_prior.py - classes and functions implementing the prior over policies

The main implementation of the POTMMCP algorithm is contained in the potmmcp/tree/policy.py file.

experiments

This directory contains scripts for running the experiments in each environment as well as Jupyter notebooks for analysing the results and the actual results files.

Results

If you run any of the experiment scripts, by default experiment results are saved to the ~/potmmcp_results directory.

Questions or Issues

If you have any questions or issues please email jonathon.schwartz@anu.edu.au or create an issue in the issue section on github.

Authors

• Jonathon Schwartz (primary author)
• Hanna Kurniwati
• Marcus Hutter

Please Cite

If you use the code in this repository or the POTMMCP algorithm, consider citing:

@article{schwartz2023combining,
  title = {Combining a Meta-Policy and Monte-Carlo Planning for Scalable Type-Based Reasoning in Partially Observable Environments},
  author = {Jonathon Schwartz and Hanna Kurniawati and Marcus Hutter},
  year = {2023},
  journal = {arXiv preprint arXiv:2306.06067},
  eprint = {2306.06067},
  eprinttype = {arxiv},
  archivePrefix = {arXiv}
}