Compositional Game Theory for Neural Networks

This repository explores the application of compositional game theory, as introduced in the paper Compositional Game Theory [1] to the analysis and enhancement of neural networks. We represent neural network components as players in open games, aiming to leverage game-theoretic tools for improved training and understanding

Repository includes:

• cgtnnlib, a library for performing the research
• data directory with some of the data we use
• doc directory with documentation
• notebooks for running experiments

As of now, the main branch is in flux. Don't expect it to be stable. Most results/revisions are available at the releases page. Older releases are at Yandex Disk.

How to run

1. Create a virtual environment
2. Install dependencies: pip install -r requirements.txt
3. Open a notebook *.ipynb file with any .ipynb reader available to you and run

cgtnnlib

The library consists of classes (with filenames beginning with capital letter) that represent problem domain (Dataset, Report, etc.) and several procedural modules:

• common.py: main functions and evaluation
• analyze.py: reads report JSON files and plots graphs
• datasets.py: dataset definitions
• plt_extras.py: matplotlib extensions
• torch_device.py: abstracts away PyTorch device selection
• training.py: training procedures
• etc.

The nn subdirectory contains PyTorch modules and functions that represent neural architectures we evaluate

The doc subdirectory contains info about datasets.

Reports

Trained models are stored in the pth/ directory. Along with each model, a corresponding JSON file is also created which contains properties like:

• started: date of report creation
• saved: date of last update
• model: model parameters, like classname and hyperparameter value
• dataset: dataset info, including the type of learning task (regression/classification)
• loss: an array of loss values during each iteration of training, for analyzing loss curves
• eval: an object that contains various values of "noise_factor", that represents noise mixed into the input during evaluation, and their corresponding evaluation metrics values: "r2" and "mse" for regression, and "f1", "accuracy", "roc_auc" for classification
• other, experiment-specific keys

Typically a report is created during the model creation and initial training, and then updated during evaluation. This two-step process creates the complete report to be analyzed by analyze.py.

References

1. N. Ghani, J. Hedges, V. Winschel, and P. Zahn. Compositional game theory. Mar 2016. https://arxiv.org/abs/1603.04641