Learning action-oriented models through active inference

This repository extends the paper Learning action-oriented models through active inference to work with more complex environments and dynamics. Animations and visualizations are added to showcase how an Active Infernece agent learn and behave in real-time.

The core code can be found in src/core, whereas scripts for running experiments can be found un src/scripts

src/data stores the data from the latest experiments/code.

src/data_downloaded contains the data (which has extra files than the code can generate directly) from the original repo.

src/data_original_code contains the data generated by the original code without modifications.

src/figs contains figures generated by latest experiments/code.

src/figs_from_download contains figures generated by the original code.

Demo

Here's a demonstration of an Active Inference agent learning and behaving in real-time:

This animation showcases how the agent chooses its actions based on the changes of approach angle in a single target reaching task over time.

The left side of the video shows the agent's trajectory in the environment it lives in (with a target). The right side shows the value of agent's Expected Free Energy and values of its components over time.

Here's another demonstration showing the agent's continual learning process as it adapts learning to reaching multiple targets in a sequence given the changing distances and approach angles: