wmayner / autonomy

Toolbox to evaluate various measures of autonomy in small simulated agents

Autonomy

autonomy is a Python package for computing measures of autonomy on small artificial agents defined by their transition probability matrices.

If you use this code, please cite the paper:

---

Albantakis L (2021) Quantifying the autonomy of structurally diverse automata: a comparison of candidate measures. Forthcoming.

---

Installation

Set up a Python 3 virtual environment and install with

pip install "git+https://github.com/Albantakis/autonomy.git"

Most functions in autonomy require PyPhi, the Python library for computing integrated information.

Note: this software is only supported on Linux and macOS. However, if you use Windows, you can run it by using the Anaconda Python distribution and installing PyPhi with conda:

conda install -c wmayner pyphi

I suggest following PyPhi's detailed installation guide.

Documentation

Getting started

The Agent object is the main object on which computations are performed. It represents the causal model of the agent as a transition probability matrix.

Initializing an agent requires a transition probability matrix (TPM). Providing a connectivity matrix is optional. Most functions also require an activity attribute. See example agents in data folder. Finally, information about the sensors, motors, and hidden units is required for most functions. See YAML files in the autonomy/phenotypes folder.

Once an agent object is defined, a full structural analysis can be initialized with the fullStructuralAnalysis(agent) function, and likewise for a full dynamical, information-theoretical, or causal analysis. These functions output a pandas dataframe with all computed values. Note: the full causal analysis can be very time consuming.

An example notebook and agent can be found in the autonomy/examples folder.

Contact

For any questions related to this code, and to report issues, please send an email to [e-mail].