Rachid Alami (LAAS-CNRS, ANITI, France)

Models and Decisional issues for Human-Robot Joint Action

This talk will address some key decisional issues that are necessary for a cognitive and interactive robot which shares space and tasks with humans. We adopt a constructive approach based on the identification and the effective implementation of individual and collaborative skills. The system is comprehensive since it aims at dealing with a complete set of abilities articulated so that the robot controller is effectively able to conduct in a flexible and fluent manner a human-robot joint action seen as a collaborative problem solving and task achievement. These abilities include geometric reasoning and situation assessment based essentially on perspective-taking and affordances, management and exploitation of each agent (human and robot) knowledge in a separate cognitive model, human-aware task planning and interleaved execution of shared plans. We will also discuss the key issues linked to the pertinence and the acceptability by the human of the robot behaviour, and how this influence qualitatively the robot decisional, planning, control and communication processes.

Thomas Eiter (Technische Universität Wien, Austria)

A Hitchhiker's Tour Through Computational Complexity in Knowledge Representation and Reasoning

Great Moments in KR

Slides in PDF

The vision of artificial intelligence with human-level capabilities of reasoning has inspired and motivated generations of researchers, starting with John McCarthy's seminal work to develop numerous formalisms and approaches towards making it a reality. Many of these formalisms are rooted in formal logic or mathematical approaches to deal with world models at a symbolic level, allowing to take aspects such as incomplete information, uncertainty, or inconsistency into account. The undecidability of first-order logic, which often served as the basis, has spurred the search for decidable fragments in order to facilitate effective reasoning. However, mere decidability turned out to be insufficient in practice, and tractability as a paradigm for efficient solvability was fostered.

Structural complexity theory, which is concerned with problem solving under resource constraints and describing its inherent difficulty, turned out to be a valuable tool in the design of formalisms and for the analysis of reasoning tasks in them. In fact the rich landscape of complexity classes with various models of computation and resource settings, has facilitated a fine-grained analysis beyond a black-white characterization of being tractable or intractable, where NP-hardness was perceived as a kiss of death for problems in practice. In turn, problems in Knowledge Representation and Reasoning (KRR) have vitalized complexity classes that were considered to be more of academic interest, and led to the development of new notions and techniques.

In this talk, we shall address the role of computational complexity in KRR. We shall consider why complexity matters and what conclusions can be drawn from complexity results beyond mere quantitative (in terms of resource consumption) results. With a focus on selected examples, we shall review some highlights and influential results, consider developments and recent trends, and perhaps risk a glimpse into the future: while regarded almost indispensable today, may the need for complexity considerations vanish?

Mateja Jamnik (University of Cambridge, UK)

How to (Re)represent it?

To achieve efficient human computer collaboration, computers need to be able to represent information in ways that humans can understand. Picking a good representation is critical for effective communication and human learning, especially on technical topics. To select representations appropriately, AI systems must have some understanding of how humans reason and comprehend the nature of representations. In this work, we are developing the foundations for the analysis of representations for reasoning. Ultimately, our goal is to build AI systems that select representations intelligently, taking users’ preferences and abilities into account.

Marta Kwiatkowska (University of Oxford, UK)

Probabilistic model checking for strategic equilibria-based decision making

Software faults have plagued computing systems since the early days, leading to the development of methods based on mathematical logic, such as proof assistants or model checking, to ensure their correctness. The rise of AI calls for automated decision making that incorporates strategic reasoning and coordination of behaviour of multiple autonomous agents acting concurrently and in presence of uncertainty. Traditionally, game-theoretic solutions such as Nash equilibria are employed to analyse strategic interactions between multiple independent entities, but model checking tools for scenarios exhibiting concurrency, stochasticity and equilibria have been lacking.

This lecture will focus on a recent extension of probabilistic model checker PRISM-games (www.prismmodelchecker.org/games/ ), which supports quantitative reasoning and strategy synthesis for concurrent multiplayer stochastic games against temporal logic that can express coalitional, zero-sum and equilibria-based properties. Game-theoretic models arise naturally in the context of autonomous computing infrastructure, including user-centric networks, robotics and security. Using illustrative examples, this lecture will give an overview of recent progress in probabilistic model checking for stochastic games and highlight challenges and opportunities for the future.

Gary Marcus (Robust AI, USA)

Taking AI to the Next Level

For nearly half a century, AI has always seemed as if it just was beyond reach, less than two decades away. Yet "strong AI" in some ways still seems elusive. In this talk, I will give a cognitive scientist's perspective on AI. What have we learned, and what are we still struggling with? Is there anything that programmers of AI can still learn from studying the science of human cognition?

David Poole (University of British Columbia, Canada)

Lessons from three decades of research into learning and reasoning with relational probabilistic models

Making probabilistic predictions from relational data, recently under the umbrella of "statistical relational AI", has a long history, and is an active research area. The talk presents some insights that everyone should know but often don't fit into research papers. These include: What is a relational model. Triples are universal representations of relations but learning with them is difficult. How embedding-based models generalize (and what they actually learn). Why embedding based models cannot be used for predicting properties. Why ranking is often used for evaluation, but is very misleading. Why entities are not like words in embedding models. Why being Bayesian implies exchangeability, which can be exploited in lifted reasoning. Lifted reasoning relies on counting, but counting relies of know what exists and knowing identity. Why identity and existence uncertainty are tricky to get right.