Reinforcement Learning Produces Dominant Strategies for the Iterated Prisoner’s Dilemma

This paper gives description of strategies trained using reinforcement learning as well as a detailed analysis of their performance in a large tournament with 176 strategies.

This directory is structured as follows:

|--- main.tex  # Source file for the paper
|--- bibliograpy.bib  # Biblioraphy
|--- environment.yml  # Conda environment
|--- assets  # All tables, images, diagrams used in `main.tex`
|--- src
     |--- players.py  # All players used
     |--- abbreviations.py  # Abbreviations for some player names
     |--- reference_keys.csv  # Citation keys for each strategy
     |--- main.ipynb  # Notebook to obtain all `../assets`
     |--- main.py  # Main file to generate tournament data
     |--- generate_cooperation_data.py  # File to generate cooperation data
     |--- write_pbs_files.py  # Script to write pbs scheduler files
     |--- submit_ml_jobs.sh  # Auto written script to submit pbs files
     |--- pbs_files  # Automatically written files
          |--- ml-0-0-1000.pbs
          |--- ...
     |--- data  # Where data file are placed

Building the article

Building the article:

The following compiles the article using Latexmk version 4.41:

$ latexmk --xelatex main.tex

The bibliography is being built using biblatex which requires biber, that comes bundled with some installs of latex but if you are having problems you might need to run (on ubuntu, similarly for other systems):

$ sudo apt-get install biber

Contributions

• Conceived of the study: MH VK
• Conducted experiments and trained strategies: VK MH MJ GK
• Analyzed the data and analytical methods: VK MH
• Wrote the paper: VK MH NG
• Created software: MH VK MJ GK
• Axelrod Library Core Team: VK OC MH

Some original training code sources

• EvolvedLookerUp by mojones
• EvolvedANN by mojones
• PSO Gambler by GDKO