Supervisors

Suitable for

Abstract

The challenge in Out-of-Distribution Dynamics Detection (O2D3) is to test whether a given sequential control environment at test-time is the same as the one the agent was trained in. This capability is central to both AI safety, and security: A robot controller may have systematic errors at test-time, or an adversary may attack the AI agent’s sensors. In both cases, if anomalies in the test-time environment are not detected as soon as possible, potentially catastrophic outcomes may follow. In this project, we are building upon the latest advances in statistics, including doubly-robust estimation techniques, information-theoretic hypothesis testing and latest advances in out-of-distribution detection, in order to surpass the current state-of-the-art in O2D3 [1]. This project is designed to lead to publication. We are looking for a highly-motivated student with interest in theory.

[1] Linas Nasvytis, Kai Sandbrink, Jakob Foerster, Tim Franzmeyer and Christian Schroeder de Witt, Rethinking Out-of-Distribution Detection for Reinforcement Learning: Advancing Methods for Evaluation and Detection, AAMAS 2024 (Oral)