IRF Paper Artifacts
This repository provides the experimental artifacts associated with our paper on intrinsic reward learning in a multi-agent combat simulation environment.
The released artifacts are intended to support transparency and methodological reproducibility. They include the final trained checkpoints used in the reported experiments, the corresponding hyperparameter configuration files, TensorBoard training logs, and evaluation result files underlying the reported metrics.
Contents
This repository is organized according to the experimental sections reported in the paper.
1. Training Stability and Convergence
This folder contains the trained model folders used for the training stability and convergence analysis.
For each algorithm, 20 independently trained runs are provided. Each run contains:
hyperparameters.yaml: training configuration fileckpt: final trained checkpoint used for evaluationevents.out.tfevents...: TensorBoard training log
2. Complexity Growth Experiment
This folder contains the artifacts for the increasing environment complexity experiment in the 3vs3, 6vs6, and 9vs9 settings.
For each environment:
networks/ contains the trained model folders for the evaluated algorithms:
results/ contains the evaluation outputs used to compute the reported performance metrics:
vs.coma-IRF_bluevs.coma-IRF_redvs.mapoca-IRF_bluevs.mapoca-IRF_red
3. Stochasticity Growth Experiment
This folder contains the artifacts for the environmental stochasticity experiment in the 6vs6 setting under different stochasticity coefficients.
For each environment configuration (e.g., weight coefficient 1 and 3):
networks/ contains the trained model folders for:
results/ contains the evaluation outputs used to compute the reported metrics:
vs.cds-IRF_bluevs.cds-IRF_redvs.emc-IRF_bluevs.emc-IRF_red
File Description
Each trained model folder typically includes the following files:
-
hyperparameters.yaml
The configuration file specifying the training settings used for the run.
-
ckpt
The final trained checkpoint used in the reported evaluation.
This repository releases the final checkpoint for each experiment rather than intermediate checkpoints.
-
events.out.tfevents...
The TensorBoard event log generated during training.
Notes on Reproducibility
The artifacts in this repository correspond to the experiments reported in the paper.
They are provided to facilitate:
- inspection of the final trained policies,
- verification of the training configurations,
- examination of the training dynamics through TensorBoard logs,
- and reproduction of the reported evaluation metrics.
The source code for the simulation environment, learning algorithms, and evaluation scripts is publicly available in the accompanying GitHub repository:
Relation to the Paper
This repository contains the experimental artifacts supporting the main results reported in the paper, including:
- training stability and convergence analysis,
- increasing environment complexity experiments,
- environmental stochasticity experiments.
The reported metrics in the paper are derived from the evaluation result files released in this repository.
Citation
If you use these artifacts, please cite our paper:
1@article{to_be_added,
2 title={TO BE ADDED},
3 author={TO BE ADDED},
4 journal={TO BE ADDED},
5 year={TO BE ADDED}
6}
