This repository contains the official implementation of the inverse decision-making method presented in the ICLR 2025 paper

Straub, D., Niehues, T. F., Peters, J., & Rothkopf, C. A. (2025). Inverse decision-making using neural amortized Bayesian actors. ICLR 2025.

To install requirements:

python -m pip install -r requirements.txt

To generate an evaluation dataset used during training, run:

python generate_eval_data.py --cost <cost_function_name>

To train an action network, run this command:

python train.py --cost <cost_function_name>

As of now, the following cost functions are available: QuadraticCost, QuadraticCostQuadraticEffort, AsymmetricQuadratic, Linex, AbsoluteCost, AbsoluteCostQuadraticEffort, InvertedGaussian

To run inference on a simulated dataset, run:

python run_inference.py --cost <cost_function_name>

To run the script that disentangles prior and costs using multiple levels of perceptual uncertainty, run:

python disentangling.py