Jekyll2026-08-05T18:42:52+00:00/feed.xmlGuillermo A. PérezGuillermo's personal websiteAdvanced Topics in Complex Event Recognition2026-04-22T00:00:00+00:002026-04-22T00:00:00+00:00/jekyll/update/2026/04/22/talk-stijnComplex event recognition (CER) requires both expressive pattern languages and efficient evaluation over high-throughput streams. This talk presents a unified formal foundation for CER based on a declarative language with precise, compositional semantics and a principled treatment of selection strategies. Queries are compiled into an automata-based model that enables efficient processing with constant-time updates per event and output-linear delay for enumerating matches, providing strong theoretical guarantees while clarifying the design space of event pattern languages.

Building on this foundation, the talk introduces an execution engine that realizes these ideas in practice through a compact automaton representation that avoids the explosion of intermediate matches. The system supports expressive features such as time windows and partition-based correlation while maintaining the desired efficiency properties. Experimental results demonstrate substantial performance improvements over prior approaches, illustrating that CER can be both expressive and scalable in real-world streaming settings.

Speaker

Stijn Vansummeren is professor at UHasselt.

Time and Place

Wednesday 22/04/2026 at 13:45pm in M.G.006.

Registration

Participation is free, but registration is compulsory.

References and Related Reading

You may be interested in the papers the content for this talk covers:

  • A Formal Framework for Complex Event Recognition ACM TODS 46(4), 2021
  • CORE: a Complex Event Recognition Engine VLDB 2022
  • https://dl.acm.org/doi/pdf/10.1145/3093742.3095106 (short paper by Stijn, from a long time ago, covering some basic ideas)
  • https://link.springer.com/article/10.1007/s00778-019-00557-w (very long survey, but sections 1 and 2 may provide some introduction)
]]>Reliable Offline Reinforcement Learning2026-03-18T00:00:00+00:002026-03-18T00:00:00+00:00/jekyll/update/2026/03/18/offline_RL_SPISafety is a crucial concern when deploying reinforcement learning (RL) algorithms in real-world applications. Furthermore, safety has many dimensions, ranging from ensuring reasonable performance to respecting predefined constraints.

In this talk, we focus on safety from an offline perspective, where the RL agent only has access to a fixed dataset of prior trajectories, without direct interactions with the environment. Given the availability of the behavior policy responsible for data collection, the primary challenge is crafting a policy that outperforms such a behavior policy.

We will present algorithms that leverage the behavior policy to compute an improved policy with high probability and discuss how to exploit the environment’s structure to improve sample efficiency.

Speaker

Thiago D. Simão is an assistant professor at Eindhoven University of Technology.

Time and Place

Monday 18/03/2026 at 13:45pm in M.A.143

Registration

Participation is free, but registration is compulsory.

References and Related Reading

]]>From PAs to POMDPs: Making Optimal Decisions Under Uncertainty2026-03-11T00:00:00+00:002026-03-11T00:00:00+00:00/jekyll/update/2026/03/11/talk-pierreStochasticity is essential in computer systems, both to make certain computations more efficient and to provide additional modelling power. In this talk, we discuss two prominent computational models that rely on stochasticity: probabilistic automata (PAs) and partially observable Markov decision processes (POMDPs).

Introduced in 1963 by Rabin, PAs were designed to give additional modelling power to standard (non)deterministic automata. Rabin demonstrated that while adding stochasticity makes automata strictly more expressive, it comes with a tradeoff: computational problems become vastly more difficult, many being undecidable. We will review the landscape of (un)decidability results for PAs.

PAs were initially studied as a theoretical model. Yet, they saw a resurgence of interest through the more general model of POMDPs, used in AI since the 1990s. POMDPs model stochastic environments where an agent must make sequential decisions. Building on the mechanics of PAs, the agent in a POMDP also receives some partial information about the current state at every time step (hence the “partially observable” in the name). This model has applications in fields such as robotics: robots must make decisions in stochastic environments using sensors that give them only partial information about their surroundings (e.g., they cannot see around a corner). We will discuss a few challenges in designing algorithms to synthesize optimal strategies in POMDPs.

The slides used during the talk are here

Speaker

Pierre Vandenhove is research associate at UMONS.

Time and Place

Wednesday 11/03/2026 at 13:45pm in M.G.006.

Registration

Participation is free, but registration is compulsory.

References and Related Reading

  • Foundational reference on probabilistic automata: Michael O. Rabin. “Probabilistic Automata”. Information and Control 6(3): 230-245 (1963).
  • Survey on PA undecidability, with concise and elegant proofs: Nathanaël Fijalkow. “Undecidability results for probabilistic automata”. ACM SIGLOG News 4(4): 10-17 (2017).
  • Introduction to POMDPs: Leslie Pack Kaelbling, Michael L. Littman, Anthony R. Cassandra. “Planning and Acting in Partially Observable Stochastic Domains”. Artificial Intelligence 101(1-2): 99-134 (1998).
  • Example of a recent algorithmic development (involving the speaker and Prof. Guillermo A. Pérez): Marius Belly, Nathanaël Fijalkow, Hugo Gimbert, Florian Horn, Guillermo A. Pérez, Pierre Vandenhove. “Revelations: A Decidable Class of POMDPs with Omega-Regular Objectives”. AAAI 2025: 26454-26462.
]]>An introduction to computational graph theory and generation algorithms2026-03-04T00:00:00+00:002026-03-04T00:00:00+00:00/jekyll/update/2026/03/04/talk-jangoedComputers are often used in combinatorics to determine if combinatorial objects with given structural or extremal properties exist as these existence problems are often too complex to solve by hand. This is done by designing and implementing generation algorithms which construct combinatorial objects from a given class (typically avoiding the generation of isomorphic copies) and analysing the resulting objects.

In this talk we will give an introduction to computational graph theory and the design of generation algorithms in particular. We will also give concrete examples of how these generation algorithms have helped to gain new insights and solve problems in mathematics and in chemistry.

Speaker

Jan Goedgebeur is professor at the KU Leuven.

Time and Place

Wednesday 04/03/2026 at 13:45pm in M.G.006.

Registration

Participation is free, but registration is compulsory.

References and Related Reading

You may be interested in this previous presentation on a related topic.

]]>Theoretically Sound Item Similarity Measures for Recommender Systems2026-02-25T00:00:00+00:002026-02-25T00:00:00+00:00/jekyll/update/2026/02/25/talk-noahQuantifying the similarity between items is a fundamental challenge in building recommender systems. However, many of the standard formulas used today are based on heuristics that might work in practice but lack a clear theoretical explanation. This reliance on trial-and-error makes it difficult to understand why certain methods perform better than others or how to improve them systematically. This seminar presents a principled approach to designing and understanding similarity measures by grounding them in the formal theory of probabilistic modelling and parameter estimation.

By adopting this perspective, we show how two new similarity measures can be derived from the Bernoulli and Multinomial distributions. We will discuss the importance of probabilistic smoothing and why choosing the right probability distribution is essential for accuracy. Finally, we demonstrate how the widely used Cosine Similarity can be reframed as the solution to a specific optimization problem. This transition from intuitive shortcuts to formal theory provides a new lens through which to view classical tools and opens a path for more rigorous research in the field.

You can find the slides used during the presentation here.

Speaker

Noah Daniëls is a PhD student in our department of computer science.

Time and Place

Wednesday 25/02/2026 at 13:45pm in M.G.006.

Registration

Participation is free, but registration is compulsory.

References and Related Reading

  • https://dl.acm.org/doi/10.1145/963770.963776
  • https://dl.acm.org/doi/10.1145/3308558.3313710
  • https://www.semanticscholar.org/paper/A-comparison-of-event-models-for-naive-bayes-text-McCallum-Nigam/04ce064505b1635583fa0d9cc07cac7e9ea993cc

And optionally the Recommender systems: the textbook chapters 2 and 3, but the essentials are covered in the first paper linked above.

]]>Decision-Focused Learning (and How to Do It Quickly)2026-02-18T00:00:00+00:002026-02-18T00:00:00+00:00/jekyll/update/2026/02/18/talk-senneMachine learning models are usually trained to be as accurate as possible: the more accurate the predictions, the better the model. But in practice, accuracy isn’t always what we actually care about. After all, predictions are only useful if they help us make good decisions. Traffic forecasts help us choose better routes, and stock price predictions guide investment decisions. This is where decision-focused learning comes in. It is a growing field in machine learning that aims to train models specifically to make predictions that lead to good decisions. However, this comes with significant challenges, including non-differentiable loss functions and computationally expensive training procedures. In this talk, I will introduce the main ideas behind decision-focused learning, highlight why it’s such a promising direction, and show how our recent work makes it much faster for one particularly important family of problems: linear optimization.

You can find the slides Senne used for his talk here.

Speaker

Senne Berden is a PhD student at the KU Leuven.

Time and Place

Wednesday 18/02/2026 at 13:45pm in M.G.006

Registration

Participation is free, but registration is compulsory: please fill in this form so that we can ensure there’s sandwiches for everyone

References and Related Reading

]]>Can abstract mathematical proofs be ran as programs?2025-04-02T00:00:00+00:002025-04-02T00:00:00+00:00/jekyll/update/2025/04/02/talk-blechschmidtThe social activities of proving and programming are traditionally thought as related but distinct. Agda is a functional proof language demonstrating that they can instead be fruitfully regarded as two sides of the same coin, notwithstanding the infinities modern mathematics abounds with.

This unification allows us to give integrated developments of algorithms which are correct by construction, to appreciate mathematics from a new computational angle, and to unlock novel ways of collaborative proof engineering and proof mining.

In order to help clarify how Agda might support your research, the talk will feature a short but complete case study in formalization – implementing and verifying the correctness of insertion sort – and also discuss Agda’s current limitations.

Speaker

Ingo Blechschmidt is a guest lecturer from Augsburg (Germany). He works in applied topos theory at the intersection of algebraic geometry and constructive mathematics.

Time and Place

Wednesday 02/04/2025, any time, from home :)

Note. Ingo cannot join us in person. Instead he has recorded his talk. You are invited to watch this recording at your own pace when you see fit. You can mail Ingo and Guillermo to discuss any questions or comments may have about the topic.

]]>Randomized Load Balancing: the Queue at the Cavity at Work2025-03-26T00:00:00+00:002025-03-26T00:00:00+00:00/jekyll/update/2025/03/26/talk-bvanhoudtRandomized load balancing algorithms in large-scale computing systems have received a lot of attention in the last few years. In this seminar we illustrate that the performance of many of these algorithms can be accessed using “the queue at the cavity approach”. This approach is an approximation method that is known to yield exact results in the large-system limit in some particular cases. Apart from illustrating how this approach works on a number of algorithms, we also touch upon some open problems in this area.

Speaker

Benny Van Houdt professor and head of the department of computer science at the University of Antwerp (Belgium), where he also obtained his Phd in 2001. He has been a post-doctoral fellow of the FWO-Flanders from October 2001 until October 2007. He is currently the Editor-in-Chief of the Performance Evaluation journal (since Jan 2018), a senior associate editor of ACM ToMPECS (since 2014) and an editorial board member of Stochastic Models (since 2016). He has been a member of the editorial board of Operations Research Letters (2007-2017) and Performance Evaluation (2011-2017).

Benny is the (co)recipient of various awards including best paper awards at ACM Sigmetrics, IFIP Performance, ITC, QEST and Valuetools. He is an elected member and officer of the IFIP working group 7.3 on Computer System Modeling and has published papers in a variety of journals such as IEEE/ACM Trans. on Networking, IEEE Trans. on Information Theory, Communications, IEEE JSAC, IEEE/OSA JOCN, Performance Evaluation, QUESTA, Journal of Applied Probability, Adv. In Applied Probability, Operations Research Letters, INFORMS JOC, EJOR, Stochastic Models, Computer Networks, Naval Research Logistics, etc.

Time and Place

Wednesday 26/03/2025 at 13:45pm in M.A.143

Registration

Participation is free, but registration is compulsory. Make sure to fill in this form.

]]>Causal discovery, modeling and inference: introduction2025-03-19T00:00:00+00:002025-03-19T00:00:00+00:00/jekyll/update/2025/03/19/talk-sleyderThis presentation will introduce the basics of causal modeling and explain how it differs from probabilistic modeling. Using Judea Pearl’s ladder of causality, we will show how causal models go beyond identifying associations (level 1) to reasoning about interventions (level 2) and counterfactuals (level 3). We will cover structural causal models and causal diagrams like directed acyclic graphs (DAGs), which help answer “what-if” questions about variables in a system. We will also discuss ways to build causal models, including using domain knowledge and methods that learn from observational data, such as score-based, constraint-based, and functional causal approaches. To make these ideas concrete, we will provide an example using the LiNGAM (Linear Non-Gaussian Acyclic Model) and share some insights from our recent research on this model.

Speaker

Sarah Leyder is a PhD candidate in UAntwerp.

Time and Place

Wednesday 19/03/2025 at 13:45pm in M.A.143

Registration

Participation is free, but registration is compulsory. Make sure to fill in this form.

References and Related Reading

A general reference:

  • Matthew J Vowels, Necati Cihan Camgoz, and Richard Bowden. D’ya like DAGs? a survey on structure learning and causal discovery. ACM Computing Surveys, 55(4):1–36, 2022.
  • Shohei Shimizu. LiNGAM: Non-gaussian methods for estimating causal structures. Behaviormetrika, 41(1):65–98, 2014.
  • Sarah Leyder, Jakob Raymaekers, and Tim Verdonck. TSLiNGAM: Directlingam under heavy tails. Journal of Computational and Graphical Statistics, 2024.
  • Judea Pearl and Dana Mackenzie. The Book of Why: The New Science of Cause and Effect. 2018.
]]>Expressiveness and Generalisation in Graph Neural Networks2025-03-12T00:00:00+00:002025-03-12T00:00:00+00:00/jekyll/update/2025/03/12/talk-florisThe expressive power of graph neural networks (GNNs) has been widely analysed through their connection to the 1-dimensional Weisfeiler–Leman (1-WL) algorithm, a key tool for addressing the graph isomorphism problem. While this link has deepened our understanding of how GNNs represent complex structures, it provides limited insight into their generalisation—specifically, their ability to accurately predict on unseen data. In this talk, we delve into the relationship between GNNs’ expressive power and their generalisation capabilities, offering a unified perspective that bridges these two critical aspects of GNN performance.

The slides used during the talk.

Speaker

Floris Geerts is professor in ADREM, UAntwerp.

Time and Place

Wednesday 12/03/2025 at 13:45pm in M.A.143

Registration

Participation is free, but registration is compulsory. Make sure to fill in this form.

References and Related Reading

]]>