Tutorial Descriptions

Agent-Based Negotiation
Reyhan Aydoğan, Özyeğin University
Catholijn M. Jonker, Delft University of Technology

Abstract
In the past few decades, there has been a growing interest in agent-based negotiation, where software agents facilitate negotiation on behalf of their users and try to reach joint agreements. It has been studied extensively in e-commerce settings, but it can also be addressed more generally as a paradigm for solving coordination and cooperation problems in multiagent systems, e.g. for task allocation. In this tutorial, we will talk about the fundamentals of agent-based negotiation as well as potential research challenges in automated negotiation. Particularly, we will discuss about designing effective negotiation mechanisms/strategies, reasoning on preferences, and predicting opponent’s unknown preferences or strategies.

Topics Covered

  • Theory (1.5 hrs)
    • Motivation
    • Potential Applications
    • Negotiation Protocols
    • Bidding Strategies
    • Opponent Modeling
    • Analysis of Negotiation (i.e. both process and outcome metrics)
    • Human-Agent Negotiation
    • State of the Art Tools for Agent-based Negotiation
  • Practice (2 hrs)
    • Building a simple negotiation strategy agent in java
    • Learn how to practically test your agent against other agents in many domains
    • Have your agent run in a competition against those of the other participants.

Level and Prerequisites
Senior Bachelor, Master or PhD level. Students should be proficient in Java. Students need to bring their own laptop and have eclipse oxygen and Java version 8 installed on their laptop. The should install Genius according to the instructions in: http://ii.tudelft.nl/genius/?q=article/releases.

Tutors’ Bio
Reyhan Aydoğan is an assistant professor at Özyegin University and a guest researcher at Delft University of Technology. As a visiting scholar, she worked at MIT and NIT. Her research focuses on the modeling, development and analysis of agent technologies that integrate different aspects of intelligence such as reasoning, decision making and learning. She is applying artificial intelligence techniques such as machine learning and semantic reasoning in designing and developing agent-based decision support systems, particularly negotiation support systems and automated negotiation tools. Dr. Aydoğan is one of the main organizers of the International Automated Negotiating Agents Competition. She has co-organized Conflict Resolution in Decision Making Workshop and Agent-based Complex Automated Negotiations and served as a program committee member in reputable conferences such as AAMAS, IJCAI, and ECAI.

Catholijn Jonker is full professor of Interactive Intelligence (0.8 fte) at the Faculty of Electrical Engineering, Mathematics and Computer Science of the Delft University of Technology and full professor of Explainable Artificial Intelligence (0.2 fte) at the Leiden Institute for Advanced Computer Science of Leiden Universiteit. She chaired De Jonge Akademie (Young Academy) of the KNAW (The Royal Netherlands Society of Arts and Sciences) in 2005 and 2006, and she was a member of the same organization from 2005 to 2010. She is a member of the Koninklijke Hollandsche Maatschappij der Wetenschappen and of the Academia Europaea. She was the president of the National Network Female Professors (LNVH) in The Netherlands from September 2013 till January 2016. Catholijn is EurAI Fellow since 2015, and EurAI board member since 2016, EurAI is the European Association for Artificial Intelligence.

Multi-Agent Reinforcement Learning and Dynamics of Learning
Daan Bloembergen, Centrum Wiskunde & Informatica
Michael Kaisers, Centrum Wiskunde & Informatica

Abstract
In this tutorial we will discuss how reinforcement learning can be applied in a multi-agent setting (MARL). The presence of other learning agents gives rise to partial observability and a non-stationary environment, violating the assumptions on which traditional RL algorithms are based. These key challenges motivate the need for a new theoretical framework to study MARL. In particular, we discuss Markov games as a multi-agent extension of MDPs, as well as three approaches to learning in this extended setting: independent learning, joint action learning, and gradient based methods. Building on this theoretical foundation we present an overview of recent algorithms that have been designed specifically to deal with the non-stationarity that arises from multi-agent interactions. In addition, we discuss evolutionary game theory (EGT) as a framework within which to study multi-agent learning dynamics formally. We discuss (normal form) games, best response sets, Nash equilibria, Pareto optimality, replicator dynamics, and evolutionarily stable strategies. The replicator dynamics are formally linked to MARL and form the basis of the evolutionary framework for studying learning dynamics of MARL. Several examples of the application of this framework to study MARL are presented.

Topics Covered
Multi-agent learning : Markov games, independent learning, joint-action learning, gradient ascent approaches, dealing with non-stationarity, opponent modelling.
Multi-agent learning dynamics : normal-form games, Nash equilibrium, evolutionary game theory, replicator dynamics, evolutionarily stable strategies.

Level and Prerequisites
The tutorial is aimed at the postgraduate level, either Master or PhD. Basic knowledge of linear algebra and calculus is assumed. In-depth knowledge of the fundamentals of reinforcement learning is required; for those new to the topic we suggest to follow the tutorial on the “Fundamentals of Reinforcement Learning” at EASSS as well. The tutorial is of interest to students and early career researchers who are interested in applying reinforcement learning in a multi-agent setting, and wish to study and understand the resulting learning dynamics. In addition, the tutorial will be suited to anyone wishing to get an up-to-date overview of the field of multi-agent reinforcement learning.

Tutors’ Bio
Daan Bloembergen is a researcher at Centrum Wiskunde & Informatica (CWI), NL. He received his Ph.D. in May 2015 from Maastricht University, where he worked on the project “Analyzing Multiagent Learning” funded by the NWO (Netherlands Organisation for Scientific Research). Previously he has also worked as a postdoctoral research associate at the University of Liverpool. His current research interests are studying the learning dynamics of multi-agent systems in general and multi-agent reinforcement learning and the relation to evolutionary game theory in particular. He has published on this topic in several journals and conferences among which JAIR, AAMAS, ECAI, and AAAI.

Michael Kaisers is a researcher at Centrum Wiskunde & Informatica (CWI), NL. He has published more than 40 research articles, primarily on the intersection of artificial intelligence and game theory. His research investigates strategies and mechanisms for strategic interactions with both cooperative and competitive elements, inspired by applications in energy systems, trading in markets and negotiation. He obtained his PhD from Maastricht University in 2012, and now leads and contributes to European joint research projects on multi-agent reinforcement learning, evolutionary dynamics, and efficient flexibility allocation in smart energy systems.

A shallow introduction to deep learning for agents
Kurt Driessens, Maastricht University
Jerry Spanakis, Maastricht University

Abstract
Deep learning has made quite an impact in the machine learning community. With the results obtained by Deep Reinforcement Learning, Deep Q-Learning and specifically Alpha Go and Alpha Go Zero, deep learning is also showing its promise in the field of intelligent agents.

The impact of deep neural networks comes from their ability to automatically discover highly informative features and data representations. Each layer in a deep neural network performs a relatively simple transformation of the data representation used by the previous layer. By constraining the information bandwidth of the network, the discovered representations are forced to be robust and information intensive. Because of this, deep neural networks enable the translation of raw, high-dimensional, but structured information into a meaningful vector space.

This tutorial will provide insight into this process by explaining how deep neural networks build or discover this latent vector space. We will focus on the information flow more than on the specific learning and tuning approaches available for deep learning. We will show examples from computer vision, natural language processing and reinforcement learning that will clearly illustrate the usefulness, but also constraints of deep learning.

Topics Covered

  • Neural networks
  • Different types of Autoencoders: Sparse, Stacked, Variational
  • Convolutional neural networks
  • Recurrent neural networks, LSTMs, GRUs
  • Deep Q Networks
  • Generative Adversarial Networks

Level and Prerequisites
The tutorial is set up as an introduction to deep neural networks. Although we assume that some background in machine learning and, for example, linear or logistic regression techniques can be useful, the tutorial should be accessible for all participants. Those with experience with deep neural networks might encounter content they are already familiar with.

Tutors’ Bio
Kurt Driessens‘ main research focus lies in the field of Machine Learning. He started his academic career working on reinforcement learning, focussing on representational issues. Since then, he broadened his scope to general machine learning techniques and is currently highly interested in the automatic discovery of data representations and data set relations.

Jerry Spanakis’ research interests include the areas of text information retrieval, topic modelling, machine learning, data (text) mining and (deep) neural networks. He is involved in several interdisciplinary projects concerning the application of Machine Learning and Artificial Intelligence to challenging domains (e.g. health, arts).

Auctions and Markets: Understanding Incentives for Multi-Agent Systems in the Presence of Scarcity
Aris Filos-Ratsikas, Oxford University
Paul Tylkin, Harvard University

Abstract
Auctions and markets are classes of mechanisms for the allocation of goods and services. As more and more of the participants in these mechanisms become artificial agents, it becomes increasingly crucial to understand the implications of the design and incentive properties of these markets. In this tutorial, we will cover the theory of auctions and markets from a design perspective, and consider applications to many real-world markets. Participants will learn the fundamental theory of markets and auctions, how to think about designing markets and the incentives of agents competing for scarce resources, and about the challenges of applying this in different settings.

Topics Covered
Fundamental theory of markets and auctions:

  • Introduction to market design
    • Arrow-Debreu Markets
    • Market (Walrasian) equilibria
  • Single-item and multi-item auctions
    • Single-item auctions: First-price, Second-price, Myerson’s optimal auction
    • Multi-item auctions: the VCG mechanism, Combinatorial auctions
  • Auctions as markets and markets as auctions

Applications of markets and auctions:

  • Spectrum, pollution, and electricity auctions
  • Online auctions (including eBay and advertising auctions)
  • Two-sided markets and applications to dating and ridesharing
  • Matching markets and applications to kidney exchange
  • Financial markets
  • Deep reinforcement learning in markets
  • Open problems and future directions

Level and Prerequisites
No formal prerequisites besides some mathematical maturity, although some familiarity with game theory is helpful. Intended primarily for graduate students.

Tutors’ Bio
Aris Filos-Ratsikas is a post-doctoral Research Assistant at the University of Oxford. He received his PhD in Theoretical Computer Science in August 2015 from the University of Aarhus, and his work is primarily on social choice, assignment problems, markets and fair division.

Paul Tylkin is a PhD candidate in Computer Science at Harvard University. He is a member of the EconCS research group, where his work focuses on computational social choice, multi-agent systems, mechanism design, reinforcement learning, and value-aligned artificial intelligence.

Logics for strategic reasoning in multi-agent systems
Valentin Goranko, Stockholm University

Abstract
Formal, logic-based approaches and methods are becoming increasingly popular and important in the modeling and analysis of the behavior and performance of players and coalitions in multi-player games and, more generally, in multi-agent systems (MAS). A variety of rich logical frameworks have been introduced for capturing various aspects of MAS, incl. knowledge, communication, and abilities to achieve strategic objectives. Such logical frameworks can be applied to formal specification and verification of elaborate properties of MAS, as well as to synthesis, via constructive satisfiability testing, of strategies, agents, or entire multi-agent systems, satisfying such formal specification. While this area is already quite rich in theoretical proposals and technical results, it is still actively expanding and maturing, and has a strong potential for further developments and applications.

In this tutorial I will introduce and discuss some of the most popular and useful logics for strategic reasoning in multi-agent systems, including variations and extensions of Coalition Logic and the alternating time temporal logic ATL. I will discuss the cases with complete information and with incomplete or imperfect information; with no memory, bounded or unbounded memory of the agents. The emphasis of the tutorial will be on understanding of the languages and semantics of these logics, and on applying them to model and specify properties of MAS, and then on applying algorithmic methods for model checking to formal verification and strategy synthesis in MAS. Finally, time permitting I will discuss some applications to multi-agent planning.

Topics Covered
Session 1: Concurrent game models. Logics for strategic reasoning in one-step multi-player games with perfect information: Coalition Logic and some variations. Reasoning about long-term strategic abilities: the alternating-time temporal logic ATL. Using model checking of ATL formulae for strategy synthesis.

Session 2: Logics for strategic reasoning with incomplete and imperfect information: some extensions and variations of ATL. Modeling and synthesis of memory-based strategies. Applications to multi-agent planning.

Level and Prerequisites
Level: Introductory.
Prerequisites: Some background on (classical) formal logic; basic knowledge of modal and temporal logics; general idea of formal specification and verification, and a general appreciation of logic-based approaches to modelling and analyzing multi-agent scenarios.

Tutor’s Bio
Valentin Goranko is a professor of Logic at the Department of Philosophy of Stockholm University with over 30 years of academic research and teaching experience in Bulgaria, South Africa, Denmark and Sweden, previously at departments of mathematics and of computer science. He has given several courses and tutorials at summer schools such as ESSLLI and NASSLLI. His main area of research expertise is theory and applications of logic in computer science, AI, multiagent systems, and game theory, where he has over 100 publications in international journals and conference proceedings. He has an ongoing research project on “Dynamical multiagent systems” funded by the Swedish Research council.

Communication issues in multi-agent decision-making
Nicolas Maudet, Sorbonne University

Abstract
Communication is a distinctive feature of multi-agent systems. When agents are involved in a collective decision-making process, the way agents exchange information is governed by a protocol. The purpose of this tutorial is to show that, even if protocols can have very different communication requirements, it is often possible to characterize the “best” protocols with the help of various techniques providing communication lower bounds of the problem considered. The tutorial will mostly be guided by examples taken from different settings, eg. information spreading and aggregation, voting (selecting a winner among a set of candidates), or resource allocation (reaching a desirable allocation of resources among agents).

Topics Covered
Analysis of protocols. Basics of communication complexity. With examples from: Computational social choice, Resource allocation, Information spreading.

Level and Prerequisites
The targeted audience are Masters’ and PhD students. No specific prerequisites for the course, except undergrad computer science knowledge.

Tutor’s Bio
Nicolas Maudet is a Professor of Computer Science at LIP6, Sorbonne University, Paris. His main research interests are artificial intelligence and multiagent systems, and concern various aspects of collective (and often distributed) decision making. His current topics of interest include: computational social choice, argumentation-based reasoning and interaction; and explanation in the context of decision-aiding. He is an Associate Editor of the Journal of Autonomous Agents and Multi-Agent Systems (JAAMAS), and an Editorial Board member of the Journal of Artificial Intelligence Research (JAIR).

Fundamentals of Reinforcement Learning
Ann Nowé, Vrije Universiteit Brussel
Diederik M. Roijers, Vrije Universiteit Amsterdam

Abstract
Reinforcement learning (RL) is a machine learning technique capable of handling difficult sequential decision problems as well as control problems. Recent successes like AlphaGo, gave RL a lot of media attention. This development of powerful RL techniques has assured that RL will become an indispensable component in the industry (e.g., manufacturing, electric power systems and grid management). Additionally, RL-based solutions for applications such as (semi-)autonomous cars, socially assistive robotics, household solar storage management, also means that RL will find its way into daily human activities.

Participants will be taught the basics of single-agent reinforcement learning and the associated theoretical convergence guarantees related to Markov decision processes. In addition we will discuss recent extensions to the RL framework, including how to incorporate knowledge through reward shaping, how to handle multiple objectives simultaneously, and how to make use of temporal abstraction through skills or options.

Topics Covered
Markov decision processes, planning: value iteration and policy iteration, learning: Q-learning and SARSA (TD-methods), actor-critic, model-based methods, reward shaping and learning from demonstrations and intro to multi-criteria extensions.

Level and Prerequisites
The tutorial is aimed at the postgraduate level, either Master or PhD. Basic knowledge of linear algebra and calculus is assumed. The tutorial is of interest to students and early career researchers who are interested in building agents that can learn to optimize their behavior using reinforcement learning. In addition, the tutorial will be suited to anyone wishing to get an up-to-date overview of the field of reinforcement learning.

Tutors’ Bio
Ann Nowé is a full professor at the Vrije Universiteit Brussel, where is heading the Artificial Intelligence lab. She graduated from the University of Ghent, as a master in Mathematics with a minor in computer science. She obtained her PhD in 1994 at the Vrije Universiteit Brussel in collaboration with Queen Mary and Westfield college London, UK. Currently, she is a full professor at the Vrije Universiteit Brussel both in the Computer Science Department of the Faculty of Sciences as in the Computer Science group of the Engineering Faculty. Her research interests include (Multi-Agent ) Reinforcement Learning and Multi-criteria Reinforcement Learning (MORL). Within MARL, she focuses on the coordination of agents with limited communication, social agents learning fair policies and the relationship between Learning Automata and Evolutionary Game Theory. Within MORL she mainly looks at settings where no assumptions on the shape of the Pareto front can be made. Recently, she also investigated the link between reward shaping and learning from demonstrations in an RL context. She also applies RL to practical applications such as control, smart grids, air-traffic control, etc.

Diederik Roijers did his PhD at the University of Amsterdam on multi-objective decision-theoretic planning, after which he was a research assistant at the University of Oxford in the field of social robotics, and then a postdoctoral researcher at the Vrije Universiteit Brussel on an FWO postdoctoral fellowship on multi-objective reinforcement learning. Currently, he is an assistant professor at the Vrije Universiteit Amsterdam in Computational Intelligence. His main research interest is on planning and reinforcement learning for multi-objective decision problems, and he is the first author of the seminal survey on this topic. He has published in the field of planning and reinforcement learning at several top venues including JAIR, AAMAS, AAAI, IJCAI, ICAPS and SIGIR. Previously. He has lectured tutorials on multi-objective planning and reinforcement learning at IJCAI 2015, ICAPS 2016, and EASSS 2016, as well as the 2-day introduction on (single-objective) reinforcement learning at the ACAI summer school on RL (2017).

Introduction to Formal Argumentation Theory
Nir Oren, University of Aberdeen
Federico Cerutti, Cardiff University

Abstract
Formal argumentation theory seeks to model the human process of argument. Argument appears fundamental to human reasoning, and formal models of argument have been used to build powerful, understandable, non-monotonic reasoning methodologies, widely applied across the AI community.

In this tutorial, we will introduce students to the main topics investigated by argumentation researchers, including concepts such as abstract and structured argumentation frameworks, dialogues, uncertainty and argument strength, and describe a real-world application of argumentation research. Finally, we will discuss some of the ongoing challenges in the field.

Topics Covered

  1. Abstract argumentation
    1. Dung’s framework (20 minutes)
    2. Labellings (10 minutes)
    3. Extended frameworks (20 minutes)
  2. Structured argumentation
    1. ASPIC family of logics (30 minutes)
    2. Argument schemes (15 minutes)
  3. Dialogue
    1. Dialogue games (15 minutes)
    2. Proof dialogues (30 minutes)
  4. MAS and argumentation
    1. Decision making (20 minutes)
    2. Trust and norms (20 minutes)
  5. Applications
    1. Explanation, dealing with humans (15 minutes)
    2. CISpaces (15 minutes)

Level and Prerequisites
We assume a basic understanding of discrete maths; the tutorial is targeted at first year PhD students.

Tutors’ Bio
Nir Oren is a Reader at the University of Aberdeen. He has been active within the argumentation community for around 15 years. His research interests include strategy in argumentation; dealing with uncertainty; explanation; and using argumentation for reasoning within multi-agent systems. He is a co-founder of the TAFA series of workshops on argumentation.

Federico Cerutti is Lecturer at Cardiff University. His research activity is focused mainly on supporting scientific enquiry via the means of formal argumentation. He has been one of the organisers of both COMMA 2010 and COMMA 2014, the two foremost conferences for formal argumentation. He is co-author of more than 50 peerreviewed papers, and co-editor of two books.

Epistemic Game Theory
Andrés Perea, Maastricht University

Abstract
Epistemic game theory is a modern, blooming approach to game theory where the reasoning of people (agents) is at the center stage. Before an agent reaches a decision in a game, he first reasons about the likely decisions and reasoning processes that other agents will adopt. In this tutorial I will provide an introduction to epistemic game theory, focusing on the key idea of common belief in rationality. The latter states that an agent does not only choose rationally himself, but also believes that others will choose rationally, believes that others believe that others will choose rationally, and so on. It will be discussed how common belief in rationality can be characterized by means of a recursive procedure, how it is different from Nash equilibrium, and how the idea can be extended to dynamic games in various ways.

Topics covered
Static games, idea of common belief in rationality, belief hierarchies and types, formal de…nition of common belief in rationality, recursive elimination procedure for common belief in rationality, simple belief hierarchies, Nash equilibrium, dynamic games, belief revision, backward induction reasoning, forward induction reasoning, belief hierarchies and types for dynamic games, common belief in future rationality, recursive elimination procedure for common belief in future rationality, backward induction, common strong belief in rationality, recursive elimination procedure for common strong belief in rationality.

Level and Prerequisites
The course will aim at a broad audience of master students and PhD students from many different fields, such as computer science, logic, artificial intelligence, mathematics, economics, and other areas. Since the tutorial is completely self-contained, no specific prior knowledge is required.

Tutor’s Bio
Andrés Perea have worked on many different topics in epistemic game theory, including Nash equilibrium, backward and forward induction reasoning in dynamic games, lexicographic beliefs, games with incomplete information and psychological games. In 2012 I have published a textbook on epistemic game theory, “Epistemic Game Theory: Reasoning and Choice”, which is the first textbook in the field.

Type-based Communication Correctness in Multi-agent Systems
Jorge A. Pérez, University of Groningen

Abstract
In many multi-agent systems, concurrency and communication are essential, dominating features. Programmers must routinely write code that involves the interaction of distributed agents, such as Web applications and cloud-based services; to avoid insidious and costly software bugs, it is critical that such interactions precisely follow communication structures, or protocols. These protocols are naturally concurrent, and therefore intrinsically complex; this calls for rigorous approaches for validating communicating agents/programs against protocol specifications.

Born in the research communities of programming languages and concurrency theory, session type systems (or simply session types) have consolidated as a rigorous, paradigm-independent methodology for enforcing correct communicating agents/programs, supported by solid formal foundations and available to practitioners in the form of verification tools and languages. Research on session types has been recently awarded with the Most Influential Paper Award presented at POPL, the Symposium on Principles of Programming Languages – see http://www.sigplan.org/Awards/POPL/.

The tutorial will cover the principles of session types, including their recently developed logical foundations. Advances on practical support for session type validation will also be presented.

Topics covered

  1. Background on type systems and models of concurrency (in particular, process calculi)
  2. Session types for binary (two-party) and multiparty protocols: definitions, properties, examples
  3. Logical foundations for concurrency
    1. CHoCo: the Curry-Howard correspondence for binary session types
    2. Analyzing multiparty protocols using CHoCo
  4. Session types for run-time verification
  5. Session types in practice: Overview of tools and languages incorporating session types
  6. Open research problems

Level and Prerequisites
The tutorial is aimed at MSc/PhD students and researchers; basic knowledge of logic and programming languages is assumed. Familiarity with programming languages and or type systems is an advantage but certainly not a prerequisite: to make the presentation self-contained, the tutorial will start by introducing and motivating all the essential concepts.

Tutor’s Bio
Jorge A. Pérez is an assistant professor (tenure track) at the University of Groningen (The Netherlands) since April, 2014. He works at the Johann Bernoulli Institute of Mathematics and Computer Science, in the group “Fundamental Computing” (led by Prof. Gerard Renardel de Lavalette). He also holds a part-time affiliation at CWI, Amsterdam (“Formal Methods” group). Jorge carries out and supervises research on models of concurrency, semantics of programming languages, and the logical foundations of concurrency. Prior to his current position, Jorge was a postdoctoral fellow at NOVA University Lisbon (2010-2014) and a PhD student at the University of Bologna (2007-2010).

Distributed Constraint Optimization for the Internet-of-Things (DCOP for IoT)
Gauthier Picard, Mines Saint-Etienne
Pierre Rust, Orange Labs

Abstract
The ever-growing nature of the Internet-of-Things (IoT) and related application domains (Smart home, Smart buildings, etc.) with numerous objects and configurations require more and more autonomy and coordination. Multi-agent systems are a suitable paradigm to model and develop applications and infrastructure to implement such IoT environments. Within the multi-agent techniques, distributed constraint reasoning is a relevant approach to model complex problems and decentralize decisions in a coordinated way. The Tutorial on Distributed Constraint Optimization for the Internet-of-Things proposes to review some DCOP solution methods relevant for the IoT context, to model a real smart home case study, and finally to program and deploy DCOP solutions methods on a real IoT environment composed of Raspberry PIs.

Topics covered
The Tutorial on Distributed Constraint Optimization for the Internet-of-Things is expected to be a half-day tutorial with a 2h-long lecture on DCOP models and algorithms and applications to the Internet-of-Things field, and a 2h-long practical work session to apply these concepts on real objects.

The first part of the tutorial (2h) will provide background on distributed constraint reasoning and optimization. We will overview the main models and solution methods (ADOPT, DPOP, Max-Sum, DSA, MGM) and discuss their numerous extensions and their applicability to IoT context. We will also discuss the Internet-of-Things application field, and the reason why DCOPs are suitable solution to equip objects will decentralized coordination schemes.

The second part of the tutorial (3h) will put into practice the aforementioned notions through a practical session during which attendants will model and deploy a DCOP solution for coordinating objects in the IoT, using dedicated Python libraries and Raspberry Pis. Before programming, attendants will work on a smart home case study before deploying the agents behavior in the objects.

The practical session relies on an Open Source library that will be available soon at https://github.com/Orange-OpenSource/pyDCOP.

Level and Prerequisites
We expect an audience interested in applying artificial agents and multiagent techniques to the Internet-of-Things, with a focus on distributed optimization.

The first part can be attended by anyone familiar with constraint-based reasoning and multi-agent systems.

Programming (e.g. Python) and problem modeling (constraint based modeling, linear programming, optimization) skills are basic requirements to follow the practical second part.

Tutors’ Bio
Gauthier Picard received a Ph.D. degree in Computer Science from the University of Toulouse in 2004, and the Habilitation degree in Computer Science from the University of Saint-Etienne in 2014. He is currently a professor in Computer Science at Mines Saint-Etienne and a full researcher at Laboratoire Hubert Curien UMR CNRS 5516. His research focuses on cooperation and adaptation in multiagent systems and distributed optimization with applications to smart grids, aircraft design, ambient intelligence, and intelligent transport. He has been in charge of the following organization activities: SASO (2016, PC Chair), SASO (2015, WS Chair), JFSMA (2015, Organization Chair), SASO (2014, Doctoral Consortium Chair), SASO (2012, Organization Chair), WI-IAT (2011, Demo Chair), ESAW’09 (2008-2009, Chair). He has also previously been member of the organization committees of the following events: ESAW (2004), JFSMA (2007),Web Intelligence Summer School (2009), EASSS (2010), MALLOW (2010), WI-IAT (2011).

Gauthier Picard teaches Artificial Intelligence, Multi-Agent Systems and DCOPs for Master-level students for more than 10 years, at Ecole des Mines de Saint-Etienne. He is also coordinating the International Master Track on Cyber-Physical and Social System. Some sample slides and practical works are available online, on DCSP and DCOP, Programming DCSP with Jason or Self-organization in Multi-Agent Systems. He also contributed on a tutorial chapter on problem solving with self-organizing multi-agent systems.

Pierre Rust is a software developer and computer science researcher (currently in PhD thesis), at Orange Labs. After more than 10 years of experience as a developer in the industry, he is now focusing on research; his main topics of interest are distributed computing, artificial intelligence and the impact of software on sustainability. He also enjoys helping fellow developers by sharing knowledge and occasionally giving lectures, notably at Mines Saint-Etienne [5], and teaches Python language at Orange.

Introduction to judgment aggregation
Daniele Porello, Free University of Bozen-Bolzano

Abstract
Judgment aggregation is a field in social choice theory and mathematical economics that studies the procedures, such as the majority rule, for aggregating sets of logical propositions. In the last ten odd years, judgment aggregation became an important support in AI, knowledge representation, and multiagent systems, as it provides a general theory for modelling situations where a central mechanism is designed to reconcile possibly conflicting Knowledge Bases, Ontologies, plans, or sets of constraints.

The focus of this tutorial is to provide the reader with a map of results of judgment aggregation for a number of logical frameworks that are used in applications to AI, multiagent systems, and knowledge representation. We start by presenting the framework of judgment aggregation. Then, we present the main results of judgment aggregation for classical propositional logic (e.g. possibility theorems, agenda characterization). Then, we extend the treatment of judgment aggregation to other logics. Firstly, we discuss a number of extensions of classical propositional logics; in particular, the family of Description Logics and a number of modal logics. Secondly, we focus on logical systems that are weaker than classical logics (non-classical logics). For instance, we present results for non-classical implications, for non-monotonic logics, and for a number of fuzzy logics.

A number of tutorial related to judgment aggregation recently presented at the major AI venues (e.g. ESSLLI, AAMAS, IJCAI). This tutorial is dedicated to highlighting the logical aspects of the theory of aggregation.

Topics covered

  1. Basics of judgment aggregation.
  2. List and Pettit’s main theorem, agenda properties, and characterisation results.
  3. Judgment aggregation in extensions of propositional logics (Description Logics, Modal logics).
  4. Judgment aggregation for the ontology integration.
  5. Judgment aggregation in non-classical weak logic (relevant, linear, and fuzzy logics).
  6. Modeling examples of logics for judgment aggregation in MAS scenarios.

Level and Prerequisites
This course is intended as an introductory course to judgment aggregation. Some basic understanding of logic will be required. No prior requirement of non-classical logics is presupposed.

Tutor’s Bio
Daniele Porello is an assistant professor at the KRDB research center of the Free University of Bozen-Bolzano. He published a number of articles on judgment aggregation, in particular related the use of non-classical logics for modelling constructive judgments and on the logics of collective agency. Moreover, he is working on logical languages for multiagent resource allocation, on logics for Multiagent Systems, and on Ontologies and Knowledge Representation.

Logics of agency
Daniele Porello, Free University of Bozen-Bolzano
Nicolas Troquard, Free University of Bozen-Bolzano

Abstract
The concept of action is an all-around topic of scholarly investigation. The formal aspects have long been studied by philosophers. It is also a fundamental concept in multiagent systems (MAS) whose main objective is the design of artificial agents who have to interact in an environment, by selecting a given course of action on the basis of their beliefs and preferences. This course is to provide a gentle overview to the logics if agency and of their applications, by focusing on logics of agency that abstract away from the process of action and concentrate on the result of the action. It includes Belnap and colleagues Seeing-To-It-That (STIT), and the logics of Bringing-It-About (BIAT) due to Kanger, Pörn, and others. They are modal logics where for each agent i there is a modality DOES_ i where DOES_ i A means that “i sees to it / brings about that A is the case”. These modalities are versatile and can be usefully combined with other notions like time, obligations, beliefs. They find many natural applications in modeling and reasoning about processes, individual agency and collective agency.

Topics covered

  1. Agency in philosophy and language (Goldman; Davidson; Belnap and Perloff’s linguistic agenda).
  2. Actions in computer science (dynamic logics).
  3. Seeing To It That (models; deliberative theories; strategies).
  4. Bringing It About.
  5. From states of affairs to resources.
  6. Modeling examples: functions of technical artefacts; plans; multiagent negotiation, bringing about resources; collective agency.

Level and Prerequisites
This course is intended as an introductory course that gives a gentle overview of the concepts of action and agency in logic. Some basic understanding of logic will be required.

Tutors’ Bio
Daniele Porello is an assistant professor at the KRDB research center of the Free University of Bozen-Bolzano. He published a number of articles on judgment aggregation, in particular related the use of non-classical logics for modelling constructive judgments and on the logics of collective agency. Moreover, he is working on logical languages for multiagent resource allocation, on logics for Multiagent Systems, and on Ontologies and Knowledge Representation.

Nicolas Troquard has taught the course “Logics of Agency” at ESSLLI 2016, and co-taught the course “Logics of Agency and MultiAgent systems” at ESSLLI 2007. From 2014 to 2016, he was a teacher in Computer Science at the University of Paris-Est Creteil. He has written extensively about logics of agency. He has co-authored articles about action and agency for the Stanford Encyclopedia of Philosophy and for the Handbook of Formal Philosophy (Springer, to appear). He gave an invited tutorial on “Logical Models of Action and Agency” at the 9th International Workshop on Argumentation in Multi-Agent Systems (ArgMAS 2012). He gave an invited lecture on logics for the resource-sensitive use of technical artefacts at the 15th International Workshop on Computational Logic in Multi-Agent Systems (CLIMA 2014).

Stable Matchings: in Theory and in Practice
Baharak Rastegari, University of Bristol

Abstract
Matching problems occur in various applications and scenarios such as the assignment of children to schools, junior doctors to hospitals, kidney transplant patients to donors and so on. In all the aforementioned, and similar, problems, it is understood that participants (which we will refer to as agents) have preferences over other agents, and the task is to find a matching (an assignment of the agents to one another) that is in some sense optimal with respect to these preferences. In this tutorial I will focus on matching markets in which all agents have preferences (as opposed to the assignment problem in which only agents on one side of a two-sided market have preferences). It is widely accepted and understood that a good and reasonable solution to such a matching problem must be stable. Intuitively speaking, a stable solution guarantees that no subset of agents find it in their best interest to leave the prescribed solution and seek an assignment amongst themselves. In this tutorial, I will give an overview of the fundamental concepts and important results in stable matching literature. I will then move on to issues and concerns that arise from practical applications, such as the presence of couples with joint preferences, social stability, and elicitation of preferences. In this tutorial you will see some polynomial-time algorithms, some hardness results, an approximation algorithms, and a quick glimpse at some techniques to circumvent hardness results such as Integer Programming and Fixed-Parameter Tractable algorithms.

Topics covered

  • Stable Marriage problem (SM) and its extensions: Stable Marriage with Ties (SMT), with Incomplete lists (SMI), or both (SMTI)
  • Hospitals-Residents problem (HR), Hospital-Residents problem with Ties (HRT), Hospital-Residents with Couples (HRC)
  • Computational results and approximation algorithms
  • Incentives and strategyproofness
  • Social stability
  • Preference elicitation
  • Integer Programming and Fixed-Parameter Tractability for stable matching problems
  • Real-life applications: hospitals-residents problem, school choice, kidney exchange

Level and Prerequisites
This tutorial is most suitable for graduate students, as well as undergraduates with a computer science background and with some familiarity with algorithm design, computational complexity and combinatorics. Some basic understanding of the main concepts in game theory is ideal, but not required.

Tutor’s Bio
Baharak (Bahar) Rastegari is a lecturer in the department of Computer Science, University of Bristol. She received his PhD in computer science in 2013 from the University of British Columbia, Vancouver, Canada. She was a post-doctoral research associate for 3 years in the University of Glasgow, working with David Manlove. Her work is primarily on matching markets and social choice.

Predicting Human Decision-Making: From Prediction to Action
Ariel Rosenfeld, Weizmann Institute of Science

Abstract
This 4-hour tutorial is intended for students and researchers who wish to broaden their knowledge and become familiar with the challenges, algorithms and empirical methodologies for predicting human decision-making. Ideally, the tutorial will allow students and researchers from different disciplines within the artificial intelligence community to enrich their current human interaction methodologies and incorporate new techniques that will better fit their discipline specific human-agent interaction needs. In the course of the tutorial I will present techniques and ideas using machine learning, game-theoretical and general AI concepts. The basis for these concepts will be covered as part of the tutorial, however a basic familiarity with the above concepts is encouraged.

The tutorial follows the AAAI 2017 and the EASSS 2017 tutorials given by the tutor (under the same tutorial name). The tutorial also follows the tutors recent book (same title) published by Morgan and Claypool, which will be available to attendees. The tutorial web page, which includes the tutorial materials along with excerpts from the book, is available at https://sites.google.com/view/predicting-human-dm.

Topics covered

  1. Motivation and examples: Agents for human rehabilitation, human-robot interaction, automated advice provision, argumentation, security, negotiation etc.
  2. The basics of human decision-making: decision theory and game theory, bounded rationality, historical and contemporary computational models of human behavior from behavioral economics, cognitive psychology and AI. Experimental evidence for human decision-making and behavior and how they are different from what we usually consider to be rational.
  3. Tools of the trade: Normative vs. descriptive approaches for predicting human decision-making, computational models and the integration of normative theories from different disciplines (e.g., social science and economics) to enhance classic prediction methods.
  4. From prediction to action: combining human decision-making prediction methods in the design of intelligent agents that interact proficiently with people. Frameworks, methodologies and applications to security, games, argumentation, advice provision, rehabilitation, human-robot interaction, personal assistants and negotiation.
  5. Additional topics and challenges: implicit vs. explicit interaction settings, enhancing prediction capabilities using additional modalities (e.g., facial expressions), transfer learning of decision policies across domains and people, the complexity of acquiring (reliable) human data, minority cases in human-generated data.

Tutor’s Bio
Ariel Rosenfeld is a Koshland Postdoctoral Fellow at the Computer Science & Applied Mathematics Department, Weizmann Institute of Science, Israel and an adjunct lecturer at Bar-Ilan University (Israel). He obtained a PhD in Computer Science at Bar-Ilan University following a BSc in Computer Science and Economics, graduated ‘magna cum laude’ from Tel-Aviv University, Israel. Rosenfeld’s research focus is Human-Agent Interaction and he has published on the topic at top venues such as IJCAI, AAAI, AAMAS and ECAI. Rosenfeld has a rich lecturing background which spans over a decade and has recently been awarded the “Best Lecturer” award for outstanding teaching by Bar-Ilan University.

Models of the Collective Behaviour of Autonomous Cars
László Z. Varga, ELTE Eötvös Loránd University

Abstract
The tutorial will explain in an easily understandable way the game theory models for the routing problem and the online routing problem. The routing problem is a problem in the classic model of urban traffic. The novel online routing problem is a problem of the traffic generated by autonomous cars. Autonomous cars continuously adapt to the real-time traffic conditions. Continuous adaptation is made available by online systems like Google Maps, Waze, TomTom, and many other navigation software that exploit real-time traffic information.

The classic traffic engineering concepts are in line with the classic game theory models. Traffic engineers assume that the traffic is distributed in accordance with the equilibrium. Algorithmic game theory investigated the routing problem and achieved important results. One of the results is the guarantee on the existence of equilibrium. Another result is the upper limit on the price of anarchy. An important result of the investigations of the evolutionary dynamics of the repeated routing games is the convergence to the equilibrium.

With the advent of the navigation systems that exploit real-time traffic information, the equilibrium assumption needs to be reinterpreted. This issue will be more critical when the traffic will be generated by autonomous cars, because autonomous cars will make informed and algorithmic decisions. The collective of autonomous cars is expected to generate almost optimal traffic. The collective behaviour of autonomous cars is modelled with the online routing game model. The online routing game model shows that non-cooperative autonomous adaptation to the real-time traffic conditions cannot guarantee optimal behaviour. The conjecture is that intention aware adaptation with a constraint on simultaneous decision making has the potential to avoid unwanted collective behaviour and to foster the convergence to the equilibrium.

Topics covered
Basic concepts of the routing problem in the traffic engineering domain (routing problem, preference, optimum, equilibrium, Braess paradox).

Classic game theory model (routing game, existence of optimum, price of anarchy, upper bound on the price of anarchy, evolutionary dynamics of repeated games, convergence to the equilibrium).

Basic concepts of the traffic generated by autonomous cars (real-time traffic information, supporting software applications, online routing problem, evolutionary dynamics in the online routing problem).

Online routing game model (online routing game, the issue of equilibrium, benefit of online data, intention awareness, and benefit in intention aware online routing games).

Level and Prerequisites
The tutorial is suitable for students with some basic understanding of the main concepts in game theory, but it is not a hard constraint. The tutorial is addressed to Master students and Ph.D. students.

Tutor’s Bio
László Z. Varga is a habilitated docent at the Faculty of Informatics of the ELTE University. He started research work in the eighties at KFKI (Budapest). He was visiting researcher in the early nineties at CERN (Geneva) and at the Department of Electronic Engineering at Queen Mary & Westfield College (University of London) where his research focused on basic and applied research into the development of multi-agent systems. Later he headed a unit at MTA SZTAKI (Budapest) researching and developing distributed internet applications. His current research interests include exploitation of real-time data in decentralized adaptive systems, like connected cars or internet of things.

Learning Agent Systems for a Reliable, Pro-Active Power Grid Management — Towards the AI-empowered Grid
Eric MSP Veith, OFFIS—Institute for Computer Science e.V.

Abstract
The European Union has set an ambitious goal: until 2050, 80% of the gross power consumption should be provided by renewable energy sources. For most countries, the go-to primary energy resources that are readily available are wind and solar radiation. However, these are variable renewable energy sources, dependent on a phenomenon we cannot control: the weather. At the same time, the notion of the smart grid introduces a vast array of sensory data that was previously not available. In the face of variable power generation and consumption, de-centralized management of the grid is being called for. This already brings Multi-Agent Systems (MAS) into the management of the power grid, at the level of Virtual Power Plants, for forecasting of load and generation, or on the market level.

This lecture intends to present representative use cases of MAS in the power grid management of today or the near future. It begins with a concise introduction into the necessary models and formulae that describe the current power grid, and outlines the building blocks that make up a modern, pro-actively managed power grid. Motivated by current demand, it presents a number of use cases and potential solutions. The lecture shows how resilient and pro-active management of future smart grids is dependent on an agent’s ability to forecast and to have an extensive model of its world, i.e., the grid and its environment. To this end, the lecture’s main part focuses on deep learning approaches as a means to both, forecasting and creating a model of the agent’s environment, offers an introduction to the theoretical foundation of deep learning and then details modern deep learning architectures and its application to real-world scenarios, up to the deriving of the hyperparameters themselves using neuroevolution. The lecture concludes with a glimpse into future research.

Topics Covered

  1. The Power Grid in aNutshell (30 min).
  2. Building Blocks of a Pro-Active Power Grid Management: Use Cases (30 min).
    1. Forecasting of Power Demand and Supply
    2. Meta-Prognoses of Wind Power or Solar Radiation
    3. Models of the Power Grid
    4. Heuristics for Optimizing Power Generation
    5. Power Factor Correction at Wind parks
  3. The Agent’s View of the World: Deep Learning Approaches to Modeling (2 h 15 min).
    1. A Gentle Introduction to ANNs, RNNs, and a Short Review of History
    2. Refresher: Linear Algebra and Tensors
    3. The Surprising Efficiency of Elman Networks for Wind Speed Forecasts
    4. Market Price Prediction with Deep Perceptrons
    5. Recurrent Networks for Time Series Prediction: Forecasting Load and Power Generation
    6. SurrogateModels: Modeling the Power Grid with Neural Networks
    7. Neuroevolution: Deriving Hyper Parameters and Neural Network Architectures Algorithmically
  4. Conclusions and Outlook of Future Research (15 min).

Level and Prerequisites
Attendees are expected to have a solid background in MAS. Additionally, solid basic knowledge of calculus and linear algebra is required. However, no further knowledge about electrical engineering with regards to the power grid, in machine learning with neural networks, or even deep learning, is required.

Tutor’s Bio
Eric MSP Veith is Senior Researcher with OFFIS—Institute for Information Technology, e.V., in Oldenburg. His research interest lies in self-learning agent systems and deep-learning-powered approaches to the management of the power grid as well as deep models of the power grid. He obtained his Ph.D. from TU Bergakademie Freiberg, Germany, with his thesis titled Universal Smart Grid Agent for Distributed Power Generation Management. Since 2011, he lectures in courses teaching distributed information processing and computer networking, and has authored study booklets on select topics of distributed systems for Wihelm Büchner University.

Intelligent Search Techniques
Mark Winands, Maastricht University
Cameron Browne, Maastricht University

Abstract
Intelligent search has been core to famous game-playing agents such as Deep Blue and Alpha Go (Zero). It enables the agent to reason about its environment and plan ahead, and has therefore many real-life applications in logistics, scheduling, healthcare, serious gaming, and even chemistry.

This tutorial will provide insight how different search techniques work with single and multiple agents. It discusses different techniques that are able to give exact solutions or converge to one. It shows even how search can be used when not much knowledge is available, or when thinking time is limited.

Topics Covered

  • Single Agent Search
  • Adversarial Search
  • Monte Carlo Tree Search
  • Applications

Level and Prerequisites
The tutorial is set up as an introduction to intelligent search. Although we assume that some background in search, (i.e. they should be knowledgeable about standard data structures and their associated algorithms), the tutorial should be accessible for all participants.  Those with experience with Monte-Carlo search will encounter content they are already familiar with.

Tutors’ Bio
Mark Winands received a Ph.D. degree in Artificial Intelligence from the Department of Computer Science, Maastricht University in 2004. He has co-organized several workshops / conferences on AI & Games. He serves as the editor-in-chief of the ICGA Journal, editor of Game & Puzzle Design, and as an associate editor of IEEE Transactions on Games. He is a member of Games Technical Committee (GTC) | IEEE Computational Intelligence Society. He has been guest editor for special issues in TCIAIG and Entertainment Computing. His research interests include heuristic search, machine learning and games. He has written more than eighty scientific publications on AI and Games.

Cameron Brown received the Ph.D. degree in computer science from the Faculty of Information Technology, Queensland University of Technology (QUT), Brisbane, Qld., Australia, in 2008. He has worked for several technology companies in the past. He has designed dozens of board games, many published, and is author of several books on game technology. Dr. Browne was Canon Research Australia’s Inventor of the Year for 1998, won the QUT Dean’s Award for Outstanding Thesis of 2008, and the 2012 GECCO “Humies” award for human-competitive results in evolutionary computation. He has been awarded an ERC consolidator grant of €2 million to start up a five-year research project called the ‘Digital Ludeme Project’.

Agent-Based Modelling for Social Simulation
Neil Yorke-Smith, Delft University of Technology

Abstract
Agent-based modelling and simulation is now accepted as a valuable tool for modelling the dynamics of complex systems, and for studying complex adaptive systems. Agent-based social simulation (ABSS) is used in diverse fields such as ecology, behavioural economics, and computational social science. ABSS shares commonalities and differences from multi-agent systems. This tutorial introduces ABSS and its uses, and shares some prominent case studies in different fields. The audience will be able to explain the role and value of ABSS, give examples of some common ABSS platforms, and describe the typical ABSS modelling process. The tutorial includes hands-on experimentation with a simple simulation in the NetLogo platform.

Topics Covered

  1. Socio-technical systems; research questions in the social sciences; some notable applications of ABSS; brief historical perspective. Case study: US energy market deregulation
  2. Simulation as a scientific methodology; generative models as a research method; types and purposes of simulation; abstraction levels, emergence. Case study: epidemic modelling 3. ABSS tools: AnyLogic, GAMA, Mason, NetLogo, Repast; comparison with (general) ABM platforms. Case study (hands on in NetLogo): Epstein’s civil society
  3. Modelling process; KISS, KIDS; role of sampling and statistical tests; validation and verification. Case study: civil society continued.
  4. Summary; ABSS, MAS, AAMAS communities; recommended reading; challenges and opportunities

Level and Prerequisites
The tutorial is aimed at the postgraduate level, either masters or doctoral. Familiarity with general AI is expected. Exposure to simulation and/or agent-based modelling is helpful, but not assumed. The tutorial could be accessible to enthusiastic bachelors level students with a computer science background.

Tutor’s Bio
Neil Yorke-Smith is an Associate Professor of Algorithmics in the Faculty of Electrical Engineering, Mathematics and Computer Science at the Delft University of Technology (TU Delft). His research focuses on intelligent decision making in complex socio-technical situations, with a particular interest in agent-based methodologies and behavioural factors in automated planning and scheduling. Previously, Dr Yorke-Smith was a faculty member at the American University of Beirut, Lebanon, and a research scientist at SRI International, USA. He holds a doctorate from Imperial College London.