This is the official repository of Conformal Decision Theory: Safe Autonomous Decisions from Imperfect Predictions by Jordan Lekeufack*, Anastasios N. Angelopoulos*, Andrea Bajcsy*, Michael I. Jordan*, Jitendra Malik*

It contains implementations of conformal controllers on three synthetic and real-data applications:

• robot motion planning around humans
• automated stock trading
• robot manufacturing.

Conformal Decision Theory(CDT) is a framework for producing safe autonomous decision despice imperfect machine learning predictions. Given a family of decision functions ${D^\lambda_t, \lambda \in \mathbb{R}}$. At each timestep $t$, the agent receive an input $x_t$, and must make a decision function $D_t^{\lambda_t}$ that will output an action $u_t:= D_\lambda(x_t)$, then incurs a loss $\mathcal{L}(u_t, y_t)$. CDT provides a way to select $\lambda_t$, such that the empirical risk is controlled: $$\frac{1}{T} \sum_{t=1}^T \mathcal{L}(D_t^{\lambda_t}(x_t), y_t) \leq \varepsilon + O(1/T)$$

With Anaconda, create a new environment and the packages in requirements.txt

    conda create -n conformaldt python=3.9
    pip install -r requirements.txt

run_factory_example.py runs the Factory example presented in the paper. run_trading_example runs the trading example.

To run the SDD example, You first need to download the dataset from the website and unzip it in your directory of choice.

    wget http://vatic2.stanford.edu/stanford_campus_dataset.zip
    unzip stanford_campus_dataset.zip

You also need to download ynet_additional_files.

Then you need to edit the default arguments to load_SDD to point to these filepaths.

You then need to create a cache for the predictions of the humans' next positions

    bash sdd/bash-cache-darts.py

Then you can create the trajectory for the robot and generate the video

    bash bash-traj.sh

The videos will be stored in sdd/videos and the results in sdd/metrics.

@article{lekeufack2024decision,
  author    = {Lekeufack, Jordan, and Angelopoulos, Anastasios N, and Bajcsy, Andrea, and Jordan, Michael I., and Malik, Jitendra},
  title     = {Conformal Decision Theory: Safe Autonomous Decisions Without Distributions},
  journal   = {arXiv},
  year      = {2024},
}