OXCAV - The Oxford Control & Verification group - is part of the Department of Computer Science at the University of Oxford, and is led by Prof. Alessandro Abate.
We tackle scientific challenges in the area of safe AI and assured autonomy. Technically, our research interests lie in the formal verification and optimal control of heterogeneous and complex dynamical models, built from first principles or learnt from data. We blend in techniques from machine learning and AI, such as Bayesian inference, RL, and game theory.
Highlights of our work are the analysis of stochastic hybrid systems, work on optimal and certified sequential decision making, and applications ranging from cyber-physical systems (smart energy and safety-critical autonomy), to modelling for the life sciences (systems biology).
See our Projects page for a list of ongoing and recent research initiatives.
A few resources for research (from published or presented material), and for teaching (from courses and workshops) are here.
We are keen to perform open science, with processes that are repeatable, reproducible, and replicable. In our work, this translates to developing and sharing data and code, and to developing software tools, most of which are available on git (or similar platforms), and often packaged and published as 'Tool Papers' - please see here.
News
The article ``Games for AI-Control: Models of Safety Evaluations of AI Deployment Protocols'' was presented at ICML24, at the Trustworthy Multi-modal Foundation Models and AI Agents (TiFA) Workshop, and is fully available at https://arxiv.org/abs/2409.07985
The article ``Data-driven abstractions via adaptive refinements and a Kantorovich metric,'' led by A. Banse and L. Romao, and presented at CDC23, has received the award from the CSS Technical Committee on Hybrid Systems.
An OXCAV paper, co-authored by Joar Skalse and Alessandro Abate, and entitled "Misspecification in Inverse Reinforcement Learning" has been selected for the Outstanding Paper Award for AAAI-23, a flagship conference in AI. This year AAAI received 8,777 submissions, of which 1,721 were accepted. Among these papers, the program committee selected the awarded paper.
The publication is openly accessible on the arXiv at: https://arxiv.org/pdf/2212.03201.pdf
We are delighted to announce that the article titled "On Imperfect Recall in Multi-agent Influence Diagrams" has won the Best Paper Award at the XIX Conference on Theoretical Aspects of Rationality and Knowledge (TARK23), held in June 2023. The contribution is led by OXCAV members J. Fox and L. Hammond, and co-authored by CS colleagues P. Harrenstein, A. Abate, and M. Wooldridge.
We are glad to report that OXCAV has three articles that will be presented at AAAI 2023. They are titled:
-
Probabilities Are Not Enough: Formal Controller Synthesis for Stochastic Dynamical Models with Epistemic Uncertainty
-
Misspecification in Inverse Reinforcement Learning
-
Low Emission Building Control with Zero-Shot Reinforcement Learning
Preprints are available on the ArXiV.
The contribution titled "All's Well That Ends Well: Avoiding Side Effects with Distance-Impact Penalties" has received a ‘best paper' award at the recent NeurIPS Workshop on “Machine Learning Safety", which was held on 9 December 2022.
The contribution investigates how the use of bespoke distance-impact metrics in the context of Reinforcement Learning, allows to prevent side effects, whilst still permitting task completion.
The publication, which is the fruit of an international collaboration across Europe and the US, is openly accessible on the arXiv at: https://arxiv.org/abs/2110.12662
We are glad to report that the article `Sampling-Based Robust Control of Autonomous Systems with Non-Gaussian Noise' will be presented at AAAI 2022. A preprint is available on the ArXiV.