AI / CI

This repository contains all code to reproduce the results of the study "Adversarial dynamics in human-machine systems" by Levin Brinkmann, Manuel Cebrian, and Niccolò Pescetelli.

Setup

These scripts have been tested with python 3.7.

The following commants

• create a new virtual environment
• update pip and install wheel
• install the main package (aci) and all its requirements
• install the djx package for parallelisation on slurm cluster

python3.7 -m venv .venv
. .venv/bin/activate
pip install --upgrade pip
pip install wheel
pip install -e ".[dev]"
pip install -e djx

Reproduce visualisations in the manuscript

Lineplots with different metrics (coordination, predictions, entropy, ...)

notebooks/plot_metrics.ipynb

Visualisations of individual episodes (color matrix)

notebooks/plot_colors.ipynb

Run on a Slurm cluster

This repository contains scripts that allows to run the simulation on slurm clusters. See below to reproduce an individual simulation locally.

First adapt the corresponding job pattern to the settings needed for your cluster: djx/djx/job_pattern/gpu.sh

Train

Run the multi-agent Q-learning simulation. This is creating a grid of different settings. A single independent slrum job is created for each point in the grid.

djx train runs/repro

Postprocess

Postprocess the traces collected during training. Here we compute metrics such as entropy and Jensen-Shannon divergence.

djx post runs/repro

Merge

Merge the post-processed files of the different runs.

merge runs/repro

Local reproduction

The following is for an AI with a central (global) architecture and a learning rate of 0.001.

Train

train runs/repro/train/rewards_individual__architecture_central__lr_0.001

Postprocess

post runs/repro/post/architecture_central__lr_0.001__rewards_individual runs/repro/train/rewards_individual__architecture_central__lr_0.001