Highway Task IRL Analysis

This repository contains the code and analysis pipeline used to reproduce the behavioral and inverse reinforcement learning (IRL) analyses for the highway task.

The code works best in a Linux environment and has been tested on:

• Ubuntu 24.04.4 LTS
• Anaconda 25.5.1

1. Install Anaconda

Install Anaconda before proceeding.

Tested version:

Anaconda 25.5.1

2. Create the Conda Environment

From the repository directory, create the Conda environment using the provided irl_env.yaml file.

Because the environment includes several large packages, increasing the pip timeout and retry limit is recommended:

export PIP_DEFAULT_TIMEOUT=600
export PIP_RETRIES=10

conda env create -f irl_env.yaml

The installation may take some time.

Activate the environment:

conda activate irl-original

Note: The provided environment is not intended to be a minimal reproduction environment. It may contain additional packages used for internal testing.

3. Register the Environment as an IPython/Jupyter Kernel

To make the Conda environment available in Jupyter or IPython:

python -m ipykernel install \
    --user \
    --name irl-original \
    --display-name "Python (irl-original)"

When running the notebooks in this repository, select:

Python (irl-original)

as the kernel.

4. Install the Analysis Package

The current directory should be the root of the cloned repository:

project_highway_irl_public/

Install the behavior_v2 package in editable mode:

pip install -e behavior_v2

Equivalently, this step can be performed by running the corresponding code block in:

1_install_env.ipynb

using the irl-original kernel.

5. Configure CUDA for TensorFlow

The analysis environment uses TensorFlow 2.7.4, which requires CUDA 11.2 and cuDNN 8.x.

Install the compatible CUDA runtime and cuDNN inside the Conda environment:

conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1

Then make the Conda environment libraries available to the runtime:

export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

This does not modify the system-wide CUDA installation.

You can verify that TensorFlow detects the GPU with:

python -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

---

Analysis Reproduction

1. AIRL Training

Run:

6_run_AIRL_data.py

Before running the script, set the sub_list variable to the subject IDs you want to analyze.

Valid subject IDs range from:

302-348

The script uses behavioral data stored in:

behavior_v2/data_v3/

and generates the corresponding AIRL models.

The version labels are internal development identifiers:

• behavior_v2: second version of the behavioral analysis environment
• data_v3: third version of the experiment, following pilot versions v1 and v2

2. Behavioral Data Analysis

Run the code blocks in:

descriptive_analysis_final.ipynb

using the irl-original kernel.

This notebook reproduces the behavioral analyses corresponding to:

• Figure 2 in the main paper
• Figure S1 in the Supplement

3. Group-Level IRL Analysis

Run the code blocks in:

IRL_group.ipynb
IRL_raw_reward.ipynb

using the irl-original kernel.

These notebooks reproduce the group-level IRL analyses corresponding to:

• Figure 3 in the main paper
• Figure S2 in the Supplement

4. IRL Trajectory Analysis

Run the code blocks in:

reward_trajectory.ipynb

using the irl-original kernel.

This notebook reproduces the IRL trajectory analyses, including:

• Figure 4 in the main paper
• Figure S3 in the Supplement