Research

My active research areas are the design and analysis of algorithms, networks, and machine learning. The threads below are how that breaks down in practice — the first two are the core programme, the third is where the methods travel, and the last two are where I follow a question because it will not leave me alone.

01 — Critical Elements In Networks

Remove one node from a communication network, a power grid or a city's road map and most of the time nothing much happens. Remove the right one and the system falls apart. My main line of work is about finding that node — and proving how much damage it can do.

Concretely: critical node and critical link detection, the most vital articulation points, maximisation of the number of connected components, minimisation of pairwise connectivity and of the largest surviving component. A recurring finding is that the classical centrality measures — degree, betweenness, closeness — are far weaker predictors of real damage than their popularity suggests, which is why exact and near-exact detection algorithms are worth their cost.

The problems are NP-hard, so the work runs on two tracks at once: exact formulations and bounds on one side, and polynomial-time algorithms with measured quality on the other.

02 — Distributed Algorithms For Multi-Hop Networks

In a wireless sensor field, a robot swarm or a UAV formation there is no central observer. Every node sees a few neighbours, messages are lost, and the topology moves while the algorithm is still running. The same connectivity questions then have to be answered from the inside.

This thread covers asynchronous distributed algorithms for edge and vertex connectivity estimation, connected and weighted dominating sets, localized vertex cover, and pairwise-connectivity minimisation — with attention to what the guarantees become when the input data are imperfect. It also reaches into applications: obstacle-avoidance path planning for autonomous mobile robots and routing for UAVs in emergency aid distribution.

03 — Machine Learning In Applied Domains

I work with people who own the data and the question, and bring the modelling. In medicine that has meant predicting the severity of acute appendicitis, mortality in burn patients, survival time in colorectal cancer, and real-time polyp detection from endoscopic images. In materials science, predicting and minimising the threshold voltage and dielectric constant of doped nematic liquid crystals. In forecasting, scrap steel prices, rail passenger volume and financial time series.

04 — Networks And The Brain

The brain is the network I find hardest to model and the most worth the trouble. With psychiatry and neuroimaging colleagues I have worked on hippocampal shape alterations in first-episode depressed patients at five-year follow-up, and in healthy young women carrying a familial risk for unipolar depression — that is, structural change that is visible before the illness is.

More recently, an fMRI study of what happens when social group support is gradually withdrawn during competition: anterior temporo-parietal junction and insula activate while the default mode network deactivates. An ongoing doctoral project in my group evaluates the therapeutic relationship in psychotherapy sessions using large language models.

05 — Questions I Follow For Their Own Sake

Two subjects sit outside my formal research areas and I keep working on them anyway, because they are the reason I wanted to do science at all.

Logical paradoxes and the limits of formal systems. The liar, Russell's paradox, Gödel incompleteness, the diagonal argument. What I find unresolved is not the technical machinery but the bookkeeping: "paradox" and "solution" are each used for several genuinely different things at once — an epistemic difficulty, the mechanism that generates it, and what a proposed response actually gives up. Separating those is a piece of work I am doing rather than merely reading about.

The double-slit experiment and quantum measurement. The experiment that will not stop being strange: what exactly counts as a record, what an observer has to be, and whether the structure of the information left behind can make a physical difference. The attraction is that these are not metaphysical questions only — some of them can be pushed onto a real quantum processor and made to answer.

These two are pursued with scientific excitement rather than a publication plan. Anything that reaches a defensible result will appear on the publications page like everything else.

© 2026 Onur Uğurlu · Google Scholar · ORCID · LinkedIn