This is a demonstration of EFE-GLean, a goal-directed active inference planning scheme as described in Active Inference with Dynamic Planning and Information Gain in Continuous Space by Inferring Low-Dimensional Latent States.

This repository contains the necessary datasets, programs, and configuration files to run the experiments and visualize the results as seen in the aforementioned paper.

• Linux on a recent x86-64 CPU (Tested on Ubuntu Linux 20.04, 22.04 and 24.04)
• Python 3.10 or 3.12

This program uses a pre-release build of LibPvrnn as its backbone. For this paper, LibPvrnn is provided as a wheel file for Python 3.10 and 3.12.

To run the experiments, it is first necessary to install and configure LibPvrnn as follows.

python -m venv .venv
pip install LibPvrnn/dist/libpvrnn-2.1a2-cp312-cp312-manylinux_2_39_x86_64.whl 
export PVRNN_SAVE_DIR="/path/to/EFE-GLean/LibPvrnn"

Set the PVRNN_SAVE_DIR environmental variable to the path to where the LibPvrnn directory is placed.

This build of LibPvrnn is not optimized for performance. It is only intended to preview the capabilities of EFE-GLean. Functionality outside of the provided scripts is not guaranteed.

For any inquiries regarding LibPvrnn, please contact the authors.

To run Experiments 1, 2 and 3 from the paper, first install LibPvrnn as shown above, then run the following commands.

pip install matplotlib
cd Tmaze
python inference_discrete_aifg.py ../LibPvrnn/configs/2d_pftdsg.toml --input_goal_reached
python inference_aifg_rgbdc.py ../LibPvrnn/configs/2d_pft_msgs_w2.toml --er_samples 100 --trials 100 --input_goal_sense --max_step_size 0.5 --max_steps 25
python inference_aifg_rgbdc.py ../LibPvrnn/configs/2d_pft_rgbdc4s.toml --er_samples 100 --trials 100 --input_goal_sense --input_corner_sense --rgb_color

To generate the figures, please use the included Jupyter notebooks in the Tmaze directory.