Invited Speakers

Ana Ozaki - University of Oslo

Knowledge Acquisition from LLMs and the Learning Theory Fjord

We look at the knowledge acquisition task by prompting large language models, where the knowledge is represented as an ontology. Within computational learning theory, one can see this task as an active learning problem, where a learner attempts to identify an ontology that reflects the knowledge of the teacher by posing queries. We consider LLMs as teachers and discuss different formats of queries tailored for learning from LLMs. We study how the expressivity of the ontology language can affect the complexity of the active learning problem. We discuss challenges and opportunities when considering LLMs as teachers and review strategies from computational learning theory to deal with fallible teachers.

David Pearce - Universidad Politécnica de Madrid

Thirty Years of Equilibrium Logic

30 years ago equilibrium logic was proposed as a general purpose, nonmonotonic formalism for knowledge representation and reasoning. A key feature was that it extended the stable model semantics for logic programs to arbitrary propositional (and later first-order) languages. it therefore provided a logical foundation for what later became known as answer set programming (ASP) as well as suggesting extensions of the ASP syntax.

Equilibrium logic is based on a well-known, nonclassical, monotonic logic, called here-and-there (HT). This makes it possible to maintain the stable model semantics while enriching the core language of ASP with new operators and thereby adding new functionalities for KR. We can do this in a uniform, systematic way benefitting from our knowledge of the underlying logic. Over the past few years several such extensions have been explored and some of them have been implemented within enriched ASP solvers. In particular there are now systems for temporal reasoning. epistemic reasoning and even deontic reasoning.

The first part of my talk will give a gentle introduction to equilibrium logic and its base logic, HT. In the second part I will try to highlight some of the more recent developments that are relevant for KR.

Gerardo Simari - Universidad Nacional del Sur in Bahia Blanca

Engineering Truly Hybrid Intelligent Systems: A Winding Two-Way Road between Perception and Reasoning

The effective management of both knowledge and data lies - to a lesser or greater extent - at the heart of any successful application of Artificial Intelligence that yields a working, performant system. Though traditionally AI has been divided into two broad approaches, in this talk we argue that the Symbolic vs. Subsymbolic divide can be more effectively seen as a spectrum between perception and reasoning. The ideas behind Neuro-symbolic (or, more generally, Hybrid) AI have been developing for decades, and are currently maturing to the point where deployed intelligent systems can take advantage of the strengths of both sides of the divide, while seeking to avoid the pitfalls of purely data-driven models. After providing a brief overview of the Hybrid AI approach, we discuss several example R&D projects that adopt it, and the key results obtained in each case. We conclude by analyzing several challenges that lie on the path to engineering hybrid intelligent socio-technical systems, that are themselves both technical and social in nature.

Son Cao Tran - New Mexico State University

Knowledge Representation meets Automated Planning: From Reasoning about Actions and Change to Planning and Model Reconciliation and Beyond

Knowledge representation and reasoning has been one of the most important research directions in AI. Research in KRR has played an important role in the development of several subareas, such as reasoning about actions and change, commonsense reasoning, automated planning, etc. In this talk, I will discuss in depth the relationship between reasoning about actions and change and automated planning and their role in recent topics such as explainable AI and epistemic planning. I will conclude with a discussion on the role of LLM in KRR and planning in general.

Sylvie Thiébaux - Australian National University & University of Toulouse

Graph Learning for Planning

State of the art methods for automated planning rely on heuristic state-space search. I will present recent work on graph representation learning to guide the search of automated planners. I will introduce graph neural network and other graph learning representations that exploit the relational structure of planning domains. They allow our planner GOOSE to learn heuristic cost estimates and state rankings from solutions to just a few small problems, and solve substantially larger problems than trained on. Perhaps surprisingly, our experimental results show that classical machine learning approaches vastly outperform deep learning on es in this context. Moreover, Greedy Best-First Search guided by our best learnt heuristics rivals with the state of the art model-based planner, Lama, on the problems of the latest International Planning Competition Learning track, leading to the possibility that learnt heuristics may replace existing model-based heuristics in the near future.