Invited talks and tutorials
Sheila McIlraith - Invited talk
Reward Machines: Formal Languages and Automata for Reinforcement Learning
Reinforcement Learning (RL) is proving to be a powerful technique for building sequential decision making systems in cases where the complexity of the underlying environment is difficult to model. Two challenges that face RL are reward specification and sample complexity. Specification of a reward function -- a mapping from state to numeric value -- can be challenging, particularly when reward-worthy behaviour is complex and temporally extended. Further, when reward is sparse, it can require millions of exploratory episodes for an RL agent to converge to a reasonable quality policy. In this talk I'll show how formal languages and automata can be used to represent complex non-Markovian reward functions. I'll present the notion of a Reward Machine, an automata-based structure that provides a normal form representation for reward functions, exposing function structure in a manner that greatly expedites learning. Finally, I'll also show how these machines can be generated via symbolic planning or learned from data, solving (deep) RL problems that otherwise could not be solved.
Alessandra Russo - Invited talk
Logic-based Machine Learning: Recent Advances and Their Role in Neuro-Symbolic AI
Learning interpretable knowledge from data is one of the main challenges of AI. There has been a growing interest in Logic-based Machine Learning, a field of Machine Learning that aims at developing algorithms and systems for learning logic-based programs that explain labelled data within the context of some given background knowledge. Logic-based machine learning has traditionally addressed the task of learning definite logic programs (with no negation). Over the last 15 years, we have made significant contributions to this field by extending it to the learning of the full class of ASP programs. Our family of learning-from-answer-set (LAS) systems boasts a number of advanced features, including the ability to support non-monotonic inductive learning, robustness to noisy examples, domain-specific optimisation and scalability with respect to large search spaces. All these features make them suitable to tackle real-world problems including learning common-sense interpretable knowledge from unstructured data. In the first part of my talk, I will present our LAS framework and show how the advances that we have made in this field make it possible to solve real-world problems in a manner that is data efficient, scalable and robust to noise. In the second part I will introduce ways in which our LAS framework can be integrated with deep learning and reinforcement learning in order to extend the applicability of such systems to problems where data are not structured.
DLV Evolution from Datalog to Ontology and Stream Reasoning
DLV has been one of the first solid and reliable integrated systems for Answer Set Programming (ASP). DLV has significantly contributed both in spreading the use of ASP and in fostering AI-based technological transfer activities.
This talk overviews the history and the recent evolution of the system, which enable effective reasoning on ontologies and streams of data, and the development of new applications.
Abstract Argumentation with Focus on Argument Claims -- An Overview
Abstract argumentation frameworks are among the best researched formalisms in the last two decades. They can be used to model discourses, provide a common ground for several nonmonotonic logics, and are employed to define semantics for more advanced argumentation formalisms. In the latter two domains, it is not the abstract argument's name, but the claim the argument represents, which should be in the focus of reasoning tasks. In this context, the fact that different arguments can represent the same claim leads to certain intricacies when it comes to the actual definition of semantics and in terms ofcomputational aspects. In this talk, we give an overview on recent results in this direction. Those include the relation between argumentation and logic programming semantics, as well as a complexity analysis of acceptance problems in terms of claims and the effect of preferences in this setting.
Termination of Reasoning with Datalog+/- Ontologies
Data-intensive applications such as ontology-based data integration and reasoning over knowledge graphs to name a few, need to exploit knowledge about the domain of the data that is available in the form of ontologies. Datalog+/- is a family of ontology languages based on Datalog that are well-suited for modeling ontologies intended for data-intensive applications. A key challenge when reasoning with Datalog+/- ontologies is non-termination. Indeed, even with very simple databases and ontologies, the inferred knowledge may be infinite due to the recursive nature of Datalog+/- rules and their ability to create new objects that are not explicitly stored in the data. This leads to the problem of uniform termination of reasoning: given a Datalog+/- ontology, is it the case that reasoning terminates for every input database? The goal of this tutorial is to illustrate the technical challenges underlying the above problem, how those challenges are affected by the mode (naive vs. smart) of reasoning, how they can be overcome, and what are the interesting open problems in this context. To this end, we are going to focus on the main paradigms that led to robust Datalog+/- ontology languages, namely guardedness and stickiness.
RML: Exploiting Logic Programming for Runtime Verification (Part I)
The tutorial is based on the most recent advances in the RML project (https://rmlatdibris.github.io/) carried out at the University of Genova, Italy. RML is a domain specific language for runtime monitoring compiled down into SWI Prolog and heavily exploiting advanced programming features as co-recursion. The tutorial will introduce runtime verification and the challenges it raises and will discuss the main features of RML and its application in a wide range of domains, from interactions in multiagent systems to robotic systems, from IoT to object-oriented interfaces.
A practical demo will complement the introduction to the language.
RML: Exploiting Logic Programming for Runtime Verification (Part II)
The tutorial is based on the most recent advances in the RML project (https://rmlatdibris.github.io/) carried out at the University of Genova, Italy. RML is a domain specific language for runtime monitoring compiled down into SWI Prolog and heavily exploiting advanced programming features as co-recursion. The tutorial will introduce runtime verification and the challenges it raises and will discuss the main features of RML and its application in a wide range of domains, from interactions in multiagent systems to robotic systems, from IoT to object-oriented interfaces.
A practical demo will complement the introduction to the language.
Answer Set Programming in the COST Action DigForASP: human rights and the problem of noise
The focus of this is Evidence Analysis, a phase where the evidence collected from digital devices is examined and aggregated to identify possible sources of proof to be presented in Court. There are weak points in evidence analysis, performed by humans, leading to undesirable uncertainty about the reliability of results. According to Kahneman (Nobel Prize) et al. "Noise: A Flaw in Human Judgment", humans seem to be unreliable decision makers, as their judgments are strongly influenced by irrelevant factors. In the Action, we advocate Artificial Intelligence and Automated (logical) Reasoning to perform evidence analysis, with a prominent role given to Answer Set Programming (ASP). This with the main objective of the protection of the rights of all the parties involved in an investigation. In this tutorial, we illustrate how ASP has inspired the Action proposal, and which developments have followed in subsequent years.