I am a Lecturer (Assistant Professor) in Machine Learning for Engineering at the School of Electrical and Electronic Engineering, University of Sheffield.

My research focuses on artificial intelligence (AI), particularly graph reinforcement learning – the intersection of reinforcement learning, planning, and machine learning on graphs. The central insight behind my work is the potential of reinforcement learning to discover novel and effective strategies for solving complex decision-making and optimisation problems, especially those that can be formulated over graphs (networks). I am motivated by the promise of AI to accelerate discovery and address important challenges in science and engineering through interdisciplinary collaborations with domain experts. The applications of my work span diverse disciplines including robotics, operations research, and the physical sciences.

News

[Oct 2026] PhD Position Available: I am recruiting a PhD student in machine learning to work on graph world models and applications in manufacturing at the University of Sheffield on a project co-supervised with Prof George Panoutsos. The position will start in October 2027 and is funded for 3.5 years. A limited number of international scholarships are available across the scheme. Please see this link for more information and to apply (deadline 13 January 2027).

[Sep 2026] In our latest work, we propose the Graph Dynamics Model (GDM), a world model designed for graph environments with stochastic dynamics and evolving topologies. We also propose the Graph Distribution Distance (GDD) metric, which addresses the thorny evaluation challenge in this setting. Check it out here.

[Sep 2026] Our work on FlowIQN will be presented at NeurIPS 2026. FlowIQN is a theoretically grounded flow-matching critic for distributional RL. By aligning flow matching with optimal transport and Wasserstein geometry, we bridge recent ideas in generative models with the foundations of distributional RL. This improves the accuracy of the return distribution and, in some cases, downstream offline RL performance.

[Aug 2026] Our paper on GNARL, which reimagines the learning of algorithm trajectories as a Markov decision process, has now been published in Transactions on Machine Learning Research (TMLR). GNARL is a very versatile method that applies to both polytime solvable and NP-hard algorithmic problems. We also released accompanying code.

For older news, see the archive.