On applications of epistemic uncertainty in decision-making: From self-driving cars to self-driving labs

Thesis

Building agents that make autonomous decisions is challenging but essential in various real-world tasks like autonomous driving, personalized healthcare, and robotics. The world they interact with is complex, governed by unknown, non-stationary and non-linear dynamics, the sensory apparatus can be faulty or noisy, and information about the world can be partial. Uncertainty plays a key role in such cases since it allows for improving the safety of the autonomous agents but also learn quickly with limited data. In this thesis, we develop tools for learning and using models based on principles of Bayesian epistemology, under which an agent infers hypotheses about the world, uses them for acting and updates them based on feedback (i.e. evidence), a process termed Bayesian action-perception loop.

In the first part of the thesis (pessimism in the face of uncertainty), we propose a method for using the model uncertainty to improve the robustness and safety of autonomous-driving agents that operate under distribution shifts. Also, we present a benchmark and methodology for assessing sequential decision-making under distribution shifts. In the second part of the thesis (optimism in the face of uncertainty), we propose methods for learning personalized treatment-effect models but also causal structures of complex dynamical systems, like the ones found in gene regulatory networks. We show how we can use epistemic uncertainty to bias the acquisition of observations and interventions (also known as experiments or actions) towards informative content that helps learn the models as quickly as possible, adapting Bayesian optimal experimental design (BOED) to the context of causal inference and discovery.

Overall, this thesis aims to advance the field of autonomous decision-making by providing practical tools for dealing with uncertainty and improving the efficiency of learning from data.

Authors: Panagiotis, T

• DOI: 10.5287/ora-5jbbox7je
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English