This repository contains supporting scripts used for the paper "Towards a General Transfer Approach for Policy-Value Networks". The scripts in this repository are only fairly small scripts, defining experiments that were run and providing links between Ludii and Polygames.
The Ludii repository contains:
- Game description files (defining the rules of all games used in this paper, interpreted by Ludii's engine)
- Code to generate tensors from Ludii's internal state action action representations, return rewards, and any other code that is required by the Python-based deep learning code. More specifically, this code is implemented in the LudiiGameWrapper.java and LudiiStateWrapper.java files.
- Code to identify which channels should map to each other for any given source-target pairing. This is implemented in the
moveTensorSourceChannels()andstateTensorSourceChannels()methods of LudiiGameWrapper.java.
The Polygames repository contains:
- All deep learning, neural networks, and MCTS code.
- JNI code for interfacing between Ludii's Java code (general game playing engine), and Polygames' C++ and Python code for search and learning.
- Script for converting trained models based on identified relations between channels.
Please use the following .bib entry to cite our paper:
@article{Soemers_2023_Transfer,
author = "D. J. N. J. Soemers and V. Mella and {\'E.} Piette and M. Stephenson and C. Browne and O. Teytaud",
journal = "Transactions on Machine Learning Research",
title = "Towards a General Transfer Approach for Policy-Value Networks",
year = "2023",
url = "\url{https://openreview.net/forum?id=vJcTm2v9Ku}"
}
See our other paper on "Deep Learning for General Game Playing with Ludii and Polygames" for more information on the bridge between Ludii and Polygames.