Introduction

This repository contains scripts to analyze game data generated with HRI-EU's fork of the level-based foraging environment (LBF). Analysis comprises the estimation of information-theoretic measures, in particular, partial information decomposition (PID), between agents' actions and the amount of food items collected.

Table of Contents

• Introduction
• Table of Contents
• Getting Started
• Estimating information-theoretic measures from LBF agent behavior
  • Running PID estimation
  • Target variable encodings for PID estimation
  • Source variable encodings for PID estimation
• Please Cite
• Related repositories
  • Contact

Getting Started

Scripts can be run within the lbf_pid_analysis subfolder. To replicate analyses shown in the paper, run get_paper_results.sh and plot_paper_figures.sh.

Running analyses presented in this repository assumes game data generated with HRI-EU's fork of the LBF environment: https://github.com/HRI-EU/lb-foraging. See the repository's README for details on running experiments.

Estimating information-theoretic measures from LBF agent behavior

Running PID estimation

To estimate PID, call

python analyze_pid_per_trial.py --folders 1 2 3 4 5 --path ../../lbf_results/experiment --target 'any_food'

where the path is the output path defined in the settings file and folders is a list of folders to compare (e.g., 3 different heuristics over one value of $c$). The target parameter is explained in detail in the next section.

Target variable encodings for PID estimation

The target parameter defines which target variable to use for PID estimation. The following variables are implemented:

• 'cooperative_actions': 1 if a food item was collected collectively, 0 otherwise
• total_food_value_collected: total food value collected in each time step, 0 otherwise (nothing collected)
• mean_food_value_collected: mean food value collected in each time step, 0 otherwise (nothing collected)
• any_food: 1 if any food item was collected by any agent in a time step, 0 otherwise (nothing collected)
• food_type: 2 if a cooperative/"heavy" food item was collected collectively in a time step, 1 if either agent collected a non-cooperative/"light" item, 0 otherwise (nothing collected)
• n_collections: number of food items collected in each step (0-2), the difference to variable 'food_type' is that this can also be two if two agents collect a food item individually in the same time step
• n_collections_agent_0: number of food items collected in each step by agent 0 (0 or 1)
• n_collections_agent_1: number of food items collected in each step by agent 1 (0 or 1)

Source variable encodings for PID estimation

Which sources are used in PID estimation and how they are encoded is specified in the settings file (config/settings.yml). There are three options for choosing source variables:

• analysis:sources: 'closest_distance': uses the distance to the closest food item in each time step
• analysis:sources: 'actions': use an agent's action in each time step
  • analysis:source_encoding: 'binary': encode actions as binary variables: 1 for action LOAD, 0 otherwise
  • analysis:source_encoding: 'individual_actions': encode individual actions from 0 = none to 5 = load

Please Cite

1. The original paper that introduces the LBF environment for reinforcement learning:

@inproceedings{christianos2020shared,
  title={Shared Experience Actor-Critic for Multi-Agent Reinforcement Learning},
  author={Christianos, Filippos and Schäfer, Lukas and Albrecht, Stefano V},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year={2020}
}

2. The paper that presents PID estimation for cooperative agent behavior:

TODO

Related repositories

• Original LBF Project Link: https://github.com/semitable/lb-foraging by Filippos Christianos ([email protected])
• HRI-EU's LBF fork that generates game data for analysis and implements additional cooperative agent heuristics: https://github.com/HRI-EU/lb-foraging by
  • Matti Krüger ([email protected])
  • Christiane Wiebel-Herboth ([email protected])
  • Patricia Wollstadt ([email protected])

Contact

Patricia Wollstadt - [email protected] - HRI-EU