My research is interdisciplinary in nature with interests spanning artificial intelligence, machine learning and data science applications across various domains. My work integrates experimental, statistical and computational approaches to solve real-world problems. Below are some key areas of my research, along with examples of my contributions and publications in each area.
1. Reinforcement Learning in Partially Observable Environments
Imagine a warehouse robot that keeps confusing indistinguishable aisles, or an automatic trading agent that must learn to act based on historical patterns (e.g., prices, volatility, macro indicators) while also considering uncertain information. My work builds reinforcement learning (RL) agents (agents that can learn via reward signals) for such messy settings, that is, when the world is only partially observable (you can’t see or are unsure about the true state) and the right action depends on past observations.
As part of this line of work, we developed three types of reward-maximising learning agents capable of dealing with different types of uncertainty (reward, state, expected and unexpected uncertainties), and showed that a simple RL mechanism that ignores state uncertainty and updates the state-action value of only the identified state accounted for how people learn better than both a Bayesian RL model that keeps track of belief states and a model that acts based on sampling from past experiences.
In a series of other projects, we introduced different RL models that are augmented with gated “working memory” (“gating” here just means deciding what to store/forget in the working memory) to be able to disambiguate aliased situations and integrate past data. Technically, I contributed two main upgrades that align with modern RL: (1) Adaptive memory for speed and accuracy – instead of fixing memory size, the agent starts small and automatically scales its working-memory capacity when accuracy stalls. (2) Uncertainty-aware learning: a Bayesian WM-based RL variant replaces hand-tuned learning rates with Kalman Temporal-Difference updates (the step size grows when estimates are uncertain and shrinks when confident) and supports Thompson-style exploration.
Selected papers
Ez-zizi, A., Farrell, S., Leslie, D., Malhotra, G., & Ludwig, C. J. (2023). Reinforcement learning under uncertainty: Expected versus unexpected uncertainty and state versus reward uncertainty. *Computational Brain & *Behavior, 6(4), 626-650. https://doi.org/10.1007/s42113-022-00165-y.
Ez-zizi, A., Farrell, S., & Leslie, D. (2015). Bayesian reinforcement learning in Markovian and non-Markovian tasks. In Proceeding of IEEE Symposium Series on Computational Intelligence, pp. 579-586. https://doi.org/10.1109/SSCI.2015.91.
2. AI for Cyber Security
How well can machine learning-based models detect entirely new variants of cyber attacks not present in the training data? Our research shows they can do it, and do it efficiently on real systems. On Windows endpoints, we built a lightweight detection system that learn from one known spyware subtype yet still can detect many unseen obfuscated families; for example, a fresh malware specimen that quietly disguises itself as a system service is detected even though that variant never appeared in training. On the network edge, we “future-proofed” intrusion detection for IoT by training a machine learning model on old attacks and correctly catching more recent held-out attacks, acting as zero-day variants.
Selected papers
Madamidola, O. A., Ngobigha, F., & Ez-zizi, A. (2025). Detecting new obfuscated malware variants: A lightweight and interpretable machine learning approach. Intelligent Systems with Applications, 25, 200472. https://doi.org/10.1016/j.iswa.2024.200472.
Adeyemi, T., Ngobigha, F., & Ez-zizi, A. (2025). Future-proofed intrusion detection for Internet of Things with machine learning. In *2025 IEEE 4th International Conference on AI in Cybersecurity (ICAIC)* (pp. 1-6). IEEE. https://doi.org/10.1109/ICAIC63015.2025.10848845.
3. Sports Performance Analytics
In this applied research strand, I work closely with colleagues in sports science to investigate the factors that influence and predict sports performance outcomes, mainly in cricket and football. In cricket, our projects analyse fielding events and sensor data from international, franchise T20 and the Hundred competition matches to understand what contextual and individual factors impact team performance. In football, our work includes modelling organised-pressing possession regains from elite women’s leagues using Bayesian Networks to uncover the tactical conditions that lead to successful attacking opportunities, as well as examining how promoted teams from elite second divisions in Europe approach and execute corner kicks.
Selected papers
Jamil, M., Manthorpe, S., MacDonald, D., & Ez-zizi, A. (2025). An examination of how fielding outcomes in international and franchise T20 and 50-over cricket are associated with bowling performances and field positions. International Journal of Performance Analysis in Sport, 1-16. https://doi.org/10.1080/24748668.2025.2455276.
Jamil, M., McEldu, M., Adshead, L., Welsh, J., Ez-zizi, A., & Beato, M. (Under revision). A spatial analysis of ball recovery locations, ball recovery methods and eventual possession outcomes in elite women’s football using a Bayesian network approach.
4. Computational Linguistics & Language Learning
Another strand of my research is in computational linguistics and cognitive modelling of language learning. I have applied machine learning and statistical techniques to language to understand how humans acquire and use language. In one study, we showed that the Rescorla–Wagner model (originally from animal learning) can account well for the computational mechanisms in a second language learning behavioural experiment. I have also contributed to research on the representation of tense and aspect in language, for example, co-developing a context-based model to explain how learners distinguish subtle rule differences like the past perfect vs. future progressive in English.
Selected papers
Ez-zizi, A., Divjak, D., & Milin, P. (2024). Error-correction mechanisms in language learning: modeling individuals. Language Learning, 74(1), 41-77. https://doi.org/10.1111/lang.12569.
Romain, L.*, Ez-zizi, A.*, Milin, P., & Divjak, D. (2022). What makes the past perfect and the future progressive? Experiential coordinates for a learnable, context-based model of tense and aspect. Cognitive Linguistics, 33(2), 251-289. https://doi.org/10.1515/cog-2021-0006. * Joint first author.
5. AI in Healthcare
My recent research focuses on developing responsible AI models for the early detection of dementia and its different variants, pioneering a new approach that enables both diagnosis of dementia and prediction of its onset over flexible time horizons. To support clinical testing and adoption, we have also implemented one of our AI models as an interactive web application, available at https://dementiaapp.aezzizi.com (login is currently password protected and can be provided upon request to clinicians). We are also building predictive models to assess the risks of behavioural and psychological symptoms of dementia, with the aim of helping families and care teams prepare in advance to reduce severe episodes and avoid hospital admissions. In parallel, I am currently leading projects in lung cancer, including models for predicting patient mortality and treatment response.
Selected papers
Ez-zizi, A., Seelam, K., Leggett, L., & Malik, B. (Under review). A flexible-horizon responsible AI model for dementia diagnosis and risk prediction in secondary care.