Autumn School on Logic and constraint Programming

Speaker: Agostino Dovier

Abstract:

Constraint Programming is declarative programming paradigm where the modeling of a problem relies on the notion of "constraint". A constraint is a restriction (precisely, a relation) imposed over the combination of values of some variables that are used to model a given problem. Solving a problem with constraints means finding a way to assign values to all its variables such that all constraints are satisfied. Constraint programming requires a modeling language and a constraint solver. Constraints can be expressed in different domains (eg, Integer Numbers, Real Numbers, Sets, Graphs,...) In the late Eighties Constraint Programming was embedded in Logic Programming, as a parametrical (wrt the constraint domain) extension of the Prolog Language (Constraint Logic Programming-CLP). This was the first available paradigm where experimenting Constraint Programming, and all modern Prolog systems offer modeling primitives and efficient constraint solvers in various domains. Since 2008, the Constraint Programming community yearly organizes a competition called Minizinc Challenge where various solvers are challenged on sets of benchmarks. In order to focus the competition on the solving part, a "standard" constraint modeling language called Minizinc that should be implemented by all the systems participating to the competition, emerged.

The tutorial will be organized as follows:

• In a live (remote) session, the basics of constraint programming will be introduced, focusing in particular on the notions of constraint propagation and constraint-based search. Global constraints and advanced search heuristics will be also briefly discussed.

• In a recorded session a practical introduction to constraint programming in Minizinc and in CLP will be presented by modeling and solving some problems.

Speaker: Daniela Inclezan

Abstract:

This presentation will introduce various applications of answer set programming (ASP) related to dynamic domains, i.e., domains characterized by change, specifically those in which change is caused by the occurrence of actions. I will start by discussing how ASP can be used to model actions and their effects in a way that addresses important problems in the field of reasoning about actions and change, such as the frame, ramification, and qualification problems. I will then show how higher-level languages that translate into ASP (so called action languages) facilitate the construction of libraries of commonsense knowledge about dynamic domains. Such libraries are fundamental for any dynamic domain application that uses ASP for reasoning purposes.

I will illustrate the use of ASP, action languages, and libraries of knowledge in a variety of applications, such as modeling theories of second language acquisition, reasoning about biological processes, demonstrating an understanding of stories via question answering, and reasoning about policy-aware intentional agents.

Speaker: Mirek Truszczynski

Abstract:

I will introduce the main formal semantics of answer-set programs, including the supported, the stable and the well-founded semantics. I will show that these semantics can be derived on top of a simple algebraic framework. I will discuss properties of the semantics and ways in which they relate to each other. I will also discuss how these semantics relate to the classical semantics of first-order theories.

To ease programming, two ideas seem to be fundamentally important: modularity and the informal semantics of answer-set programs, the latter meaning an informal way to read (understand) programs and their elements as precise natural language narrative. I will discuss both notions and argue that, when properly developed, they offer an elegant and effective methodology of writing correct and easy to understand programs.

Speakers: Alessandra Russo & Mark Law

Abstract:

Logic-based learning is a type of Machine Learning that aims at learning interpretable models of the world from observations (positive and negative), using also existing relevant knowledge and/or integrity constraints. The learned models, also called hypotheses, can explain the positive observations, classify them from negative observations, and because of their generalisation property, accurately predict unseen observations. A key characteristic of logic-based learning is that logic is used as the underlying unifying representation language for observations, background knowledge and hypotheses, which makes the learned models human interpretable. Various approaches have been developed in AI since the 1980’s and recent advances have also seen an increased application of this form of machine learning in real-world problems including privacy and security, software engineering, automata learning, language grammars, etc.

This tutorial is divided into two parts. The first part will provide an overview of the current state-of-the-art of logic-based learning, starting from its key foundation concepts and principles, and some key algorithmic aspects that are common to a variety of logic-based learning systems. This will be followed by a summary of recent real-world applications of state-of-the-art systems for logic-based learning, with an indication of open challenges and future directions. The second part will go more in-depth on approaches for learning under the answer set semantics, beginning with the traditional methods of brave and cautious learning. It will be shown how traditional logic-based learning methods are incapable of learning general answer set programs, and that a more advanced method, called Learning from Answer Sets (LAS), is required.