Multi-Modal Conformal Prediction Regions with Simple Structure by Optimizing Convex Shape Templates

Generating efficient conformal prediction regions with simple shapes.

Motivation

Conformal Prediction is a popular method for uncertainty quantification, resulting in prediction regions around a predictor that contain the true output with high probability. The shape of these regions, determined by a non-conformity function, can significantly impact the performance of downstream tasks in robotics.

Methodology

Our approach consists of four key steps:

1. Density Estimation
2. Clustering
3. Shape Construction
4. Conformalization

To learn more, check out our paper!

Key Contributions

1. A framework for generating non-conformity scores that result in efficient conformal prediction regions.
2. A Python toolbox for easy implementation of our method.
3. Applications demonstrated in F16 fighter jet and autonomous vehicle scengarios.
4. Up to 68% reduction in prediction region area compared to benchmark methods.

Usage Instructions

Our library is available on PyPi and can be installed using the following command:

pip install conformal-region-designer

To use the library, we only need three lines of code, as shown below:

import conformal_region_designer as crd
pcr = crd.ConformalRegion('kde', 'meanshift', 'ellipse', 0.90)
pcr.fit(Z_cal_one)
pcr.conformalize(Z_cal_one)

Links

• Paper
• Code