Learning Monitor Ensembles for Operational Design Domains

Hazem Torfah, Aniruddha R. Joshi, Shetal Shah, S. Akshay, Supratik Chakraborty, and Sanjit A. Seshia. Learning Monitor Ensembles for Operational Design Domains. In 23rd International Conference on Runtime Verification (RV), pp. 271–290, Lecture Notes in Computer Science 14245, Springer, 2023.

Download

[HTML] 

Abstract

We investigate the role of ensemble methods in learning runtime monitors for operational design domains of autonomous systems. An operational design domain (ODD) of a system captures the conditions under which we can trust the components of the system to maintain its safety. A runtime monitor of an ODD predicts, based on a sequence of monitorable observations, whether the system is about to exit the ODD. For black-box systems, a key challenge in learning an ODD monitor is obtaining a monitor with a high degree of accuracy. While statistical theories such as that of probably approximate learning (PAC) allow us to provide guarantees on the accuracy of a learned ODD monitor up to a certain confidence probability (by bounding the number of needed training examples), practically, there will always remain a chance, that using such a one-shot approach will result in monitors with a high misclassification rate. To address this challenge we consider well-known ensemble learning algorithms and utilize them for learning ODD ensembles. We derive theoretical bounds on the estimated misclassification risk of ensembles, showing that it reduces exponentially with the number of monitors and linearly with the risk of individual monitors. An empirical evaluation of the impact of different ensemble learning methods on a case study from autonomous driving demonstrates the advantage of this approach.

BibTeX

@inproceedings{torfah-rv23,
  author       = {Hazem Torfah and
                  Aniruddha R. Joshi and
                  Shetal Shah and
                  S. Akshay and
                  Supratik Chakraborty and
                  Sanjit A. Seshia},
  editor       = {Panagiotis Katsaros and
                  Laura Nenzi},
  title        = {Learning Monitor Ensembles for Operational Design Domains},
  booktitle    = {23rd International Conference on Runtime Verification (RV)},
  series       = {Lecture Notes in Computer Science},
  volume       = {14245},
  pages        = {271--290},
  publisher    = {Springer},
  year         = {2023},
  abstract  = {We investigate the role of ensemble methods in learning runtime monitors for operational design domains of autonomous systems. An operational design domain (ODD) of a system captures the conditions under which we can trust the components of the system to maintain its safety. A runtime monitor of an ODD predicts, based on a sequence of monitorable observations, whether the system is about to exit the ODD. For black-box systems, a key challenge in learning an ODD monitor is obtaining a monitor with a high degree of accuracy. While statistical theories such as that of probably approximate learning (PAC) allow us to provide guarantees on the accuracy of a learned ODD monitor up to a certain confidence probability (by bounding the number of needed training examples), practically, there will always remain a chance, that using such a one-shot approach will result in monitors with a high misclassification rate. To address this challenge we consider well-known ensemble learning algorithms and utilize them for learning ODD ensembles. We derive theoretical bounds on the estimated misclassification risk of ensembles, showing that it reduces exponentially with the number of monitors and linearly with the risk of individual monitors. An empirical evaluation of the impact of different ensemble learning methods on a case study from autonomous driving demonstrates the advantage of this approach.},
}