Abstract: Recent developments in AI have shown the remarkable power of deep learning, deep reinforcement learning, and LLMs. The resulting systems, however, require large amounts of data, are not transparent or reliable, and struggle in reasoning and planning tasks. In this tutorial, I'll review a top-down approach to learning, reasoning, and planning, that makes a clear distinction between what is to be learned and how, and which builds on both symbolic and neural methods. Three concrete, key learning problems in planning will be considered: learning lifted world models, learning general policies and heuristics, and learning problem decompositions.

Abstract: An ontology specifies an abstract model of a domain of interest via a formal language that is typically based on logic. Tuple-generating dependencies (tgds) and equality-generating dependencies (egds), originally introduced as a unifying framework for database integrity constraints, and later on used in data exchange and integration, are well suited for modeling ontologies that are intended for data-intensive tasks. In recent years, there has been an extensive study of tgd- and egd-based ontologies and of their applications. In those studies, model theory plays a crucial role and it typically proceeds from syntax to semantics. In other words, the syntax of an ontology language is introduced first and then the properties of the mathematical structures that satisfy ontologies of that language are explored. There is, however, a mature and growing body of research in the reverse direction, i.e., from semantics to syntax. Here, the starting point is a collection of model-theoretic properties and the goal is to determine whether or not those properties characterize some ontology language. Such results are welcome as they pinpoint the expressive power of an ontology language in terms of insightful model-theoretic properties. The goal of the lecture is to present a comprehensive overview of model-theoretic characterizations of tgd- and egd-based ontology languages that are encountered in symbolic AI.

Abstract: TBA

Abstract: In the field of knowledge representation and reasoning, rules are a convenient way of representing knowledge. But in real-life applications, every rule has its exceptions. Formally, such rules “if A then typically B” can be represented by conditionals. This course introduces reasoning with conditionals, a form of non-monotonic reasoning. Their study in AI dates back to seminal works by Reiter and McCarthy, and is still an active research field today. We will cover two core topics: (a) finding adequate semantic models for conditionals, (b) finding adequate inductive inference operators, i.e., operators determining what inferences should follow from a set of conditionals. Beginning with foundational systems and tools, such as the system P postulates and ordinal conditional functions (OCFs), we explore established and more recent inductive inference operators, including rational closure, system W, and disjunctive rational closure. Key representation results connect syntactic postulates to semantic models. Later parts of the course cover advanced properties of inference operators like syntax splitting and knowledge-based monotony. Students will gain both a historical perspective and state-of-the-art knowledge about reasoning with conditionals, enabling them to critically assess research in this area and giving them a first insight into the field of knowledge representation and reasoning.

Abstract: TBA

Abstract: Computational Argumentation (CA) is a collection of approaches for dialectical reasoning, where collections of arguments are typically evaluated ascertain their acceptability by means of semantics. Recently, CA has gained substantial traction as a research area within knowledge representation and reasoning, notably because of its potential to bridge human and machine reasoning, and symbolic (logic-based) and subsymbolic (machine learning-based) inference. This tutorial gives an overview of how CA and machine learning can be used in tandem to go beyond the boundaries of what is achievable when either of the two is used in isolation. Specifically, it focuses on: (i) how CA can be used to enhance machine learning, e.g., by improving accuracy or facilitating explainability; and (ii) how machine learning can be used for CA-based reasoning, e.g., by improving the computation time or convergence of certain semantics.

Abstract: In this tutorial, I will introduce a family of logical languages and a rule-based semantic framework for formalizing causal reasoning and causal concepts including actual causality and counterfactual conditionals. I will then present techniques for automating these languages, based on satisfiability and model checking.  Finally, I will illustrate how these languages and decision procedures can be applied to causal reasoning in the legal domain.

Abstract: Knowledge compilation studies algorithms allowing to convert between different representations of Boolean functions. This is particularly relevant when one is interested in getting a better understanding of the models of a given Boolean function. More often than not and because it is mainly the most natural way of doing it, the models of a Boolean function are implicitly defined as the one satisfying a set of constraints. The main problem with this representation is that even finding one model is NP-hard. While SAT solvers often excel at this task, they will be less efficient for tasks requiring an understanding of every model, such as when one is interesting in counting the satisfying assignments, uniformly sample a subset of them, or finding some "optimal" satisfying assignment a weight on them. On the other hands, listing every model can often be too expensive and a successful line of research has been to study data structures allowing to represent Boolean functions in a succinct yet tractable way.

In the first part of this tutorial, we will study a few such representations, explain why they are interesting and present the main algorithms used in practice to construct them. We will also showcase a few existing tools and their applications for real problems. In a second part, we will study lower bounds on such representations by exhibiting Boolean functions which cannot be efficiently represented using the data structures from the previous part.