Supporting Human Activity Performance Using

Supporting Human Activity Performance Using Argumentation-Based Technology Esteban Guerrero Rosero

Ph Licentiate Thesis Department of Computing Science Umeå University SWEDEN 2014

Supporting Human Activity Performance Using Argumentation-Based Technology

Esteban Guerrero Rosero

Licentiate Thesis

D EPARTMENT OF C OMPUTING S CIENCE U ME A˚ U NIVERSITY SWEDEN 2014

Department of Computing Science Ume˚a University SE-901 87 Ume˚a, Sweden [email protected] c 2014 by Esteban Guerrero Copyright

c Guerrero, J.C. Nieves, H. Lindgren 2014 Except Paper I, E. c Paper II, IOS Press, 2013 c Guerrero, H. Lindgren, J.C. Nieves 2013 Paper III, E.

Cover paintings: • (Front) Title: “Vendedoras de esperanza”; author: Edmundo Rosero Burgos; country: Colombia; techinque: oil on canvas. • (Back) Title: “Ancestro”; author: Edmundo Rosero Burgos; country: Colombia; techinque: oil on canvas. ISBN 978-91-7601-136-2 ISSN 0348-0542 UMINF 14.20 Printed by Print & Media, Ume˚a University, 2014

Abstract

The aim for the research presented in this thesis is to develop theories, methods and technology which can detect, represent and evaluate purposeful human activity based on information, which is inconsistent, incomplete and which includes information about an individual’s needs, goals and motives. The purpose is to provide instruments to the fields of assistive technology and ambient assisted living, which have the potentials to advance the treatment of semantic information for the purpose of making decisions and reasoning about complex human activity in daily life. Furthermore, based on sound interpretations of activity and evaluation of activity performance, tailored recommendations can be provided to the individual. This is achieved by integrating theoretical models of human activity with formal argumentation theory, and in this way create a model for representing the knowledge. Also based on the activity-theoretical models of human activity and argumentation, a computational model for common-sense reasoning was built based on the notion of fragments of activity. Moreover, a calculus model for evaluating activity performance considering the so called activity fragments was developed. These theoretical models were forming a base for implementing an argument-based reasoner engine, which manages incomplete and inconsistent information. These results were implemented in a prototype system designed for aiding an individual in improving health in daily life, for the purpose of testing the theories and methodologies with potential users. Information obtained by the sensors of a mobile phone and by assessments done by the individual regarding priorities, goals and motives were used in the analysis. Preliminary results are fed into further development, and future work includes user studies over a longer period of time. Furthermore, assistive technology for monitoring human activities in different contexts such as mining and construction industries are also being developed and tested. Future work includes also the development of methods for reasoning about activities based on data-stream sources, and taking into account that human needs, goals and motives vary over time. Moreover, methods for automatic configuration of intelligent collaborative agents and for fusing of heterogeneous data sources for providing tailored services will also be developed and evaluated in the ambient assisted living context.

iii

iv

Sammanfattning

Syftet med forskningen som presenteras i denna avhandling a¨ r att utveckla teorier, metoder och teknologi som kan detektera, representera och utv¨ardera m¨anniskors aktivitetsutf¨orande, baserat p˚a information som a¨ r inkomplett, tvetydig och som inkluderar information om individens behov, m˚al och prioriteringar. M˚alet a¨ r att utveckla nya mer avancerade instrument a¨ n vad som anv¨ands idag f¨or att st¨odja individen i dagliga aktiviteter med hj¨alp av s˚a kallade smarta-heml¨osningar eller assistansteknologi (ambient assisted living, assistive technology). De nya instrumenten har funktionalitet som inkluderar ny teknik f¨or beslutsfattande och f¨or att resonera om m¨anniskans aktivitet i dagliga livet. Baserat p˚a logiskt sunda slutsatser om p˚ag˚aende aktiviteter och utv¨ardering av aktivitetsutf¨orandet, kan personanpassade rekommendationer och st¨od ges till individen. Detta uppn˚as genom att integrera teoretiska modeller av m¨anniskans utf¨orande av aktivitet med formell argumentationsteori i en modell f¨or att representera relevant kunskap. P˚a basen av dessa teorier utvecklades a¨ ven en ber¨akningsmodell f¨or “vardagsresonemang” som kan hantera os¨aker och inkomplett information. Dessutom utvecklades en ber¨akningsmodell f¨or att utv¨ardera aktivitetsutf¨orande p˚a basen av konceptet “fragment av aktivitet”. Dessa teoretiska modeller bildar basen f¨or en argumentations-logisk exekveringsmotor som kan hantera inkomplett och inkonsistent information. Dessa resultat implementerades i en prototyp designad f¨or att st¨odja m¨anniskor f¨or att f¨orb¨attra h¨alsan i vardagen, i syfte att testa och utv¨ardera teorier, metoder och teknologin med potentiella slutanv¨andare. Information insamlad med en mobiltelefons sensorer kombinerades i analyser med information insamlat i sessioner av behovs- och m˚alformuleringar med individen. Prelimin¨ara resultat fr˚an en pilotstudie har anv¨ants i fortsatt utveckling, och kommer att ligga till grund f¨or framtida studier o¨ ver en l¨angre tidsperiod av anv¨andande. Dessutom a¨ r teknologin under anpassning f¨or fler anv¨andningsomr˚aden, som monitorering h¨alsa i gruv- och byggindustrin. Framtida forskning inkluderar a¨ ven utveckling av metoder f¨or resonerande om aktiviteter med str¨ommande data, och som hanterar variationer i m˚al, motiv och preferenser hos individen o¨ ver tid. Metoder f¨or konfigurering av team av samarbetande mjukvaruagenter i en hemmilj¨o, och hanterande av information fr˚an flera datak¨allor kommer att utvecklas och utv¨arderas i en “smart” hemmilj¨o.

v

vi

Preface

This Licentiate Thesis consists of the following papers: Paper I

E. Guerrero, J.C. Nieves and H. Lindgren. (2014). Arguing through the Well-Founded Semantics: an Argumentation Engine. Journal paper, (submitted).

Paper II

J. C. Nieves, E. Guerrero and H. Lindgren. (2013). Reasoning about Human Activities: an Argumentative Approach. In M. Jaeger, T. D. Nielsen, and P. Viappiani, editors, Twelfth Scandinavian Conference on Artificial Intelligence, SCAI 2013, Aalborg, Denmark, November 20-22, 2013, volume 257 of Frontiers in Artificial Intelligence and Applications, pages 195204. IOS Press, 2013. 1 .

Paper III

E. Guerrero, H. Lindgren and J.C. Nieves. (2013). ALI, an Ambient Assisted Living System for Supporting Behavior Change. VIII Workshop on Agents Applied in Health Care (A2HC 2013): 81-92, 2013.

In addition to the papers included in this thesis, other publications were published within the studies but not contained in this licentiate thesis. Submitted journal articles, subsume contributions of the following articles: • J. C. Nieves, E. Guerrero, J. Baskar and H. Lindgren. (2014). Deliberative Argumentation for Smart Environments. In 17th International Conference on Principles and Practice of Multi-Agent Systems (PRIMA 2014), in press. • E. Guerrero, J.C. Nieves and H. Lindgren. (2014). Towards a Human-Centric Argument-based Decision Making Framework. Introducing: ALI, an Assisted LIving System. Journal paper, (submitted). • E. Guerrero, H. Lindgren and J.C. Nieves. (2013). ALI, an Assisted Living System Based on a Human-Centric Argument-Based Decision Making Framework. In proceedings 13th Workshop on Computational Models of Natural Arguments (CMNA 2013). 46-51, 2013. • E. Guerrero, J.C. Nieves and H. Lindgren. (2013). ALI: an Assisted Living System for Persons with Mild Cognitive Impairment. In 26th IEEE International Symposium on Computer-Based Medical Systems (CBMS 2013). DOI 10.1109/CBMS.2013.6627861, 526-527, 2013. 1

Reprinted by permission of IOS Press.

vii

viii

Acknowledgments

The acknowledgments section is similar to three classic problems in Artificial Intelligence: 1) The Frame problem [McCarthy and Hayes 1969] the difficulty of indicating and inferring all those things that do not change when thanking (action) are performed and time passes, I can forget something or somebody; 2) The Ramification problem [Finger 1987] The difficulty is that it is unreasonable to explicitly record all of the consequences of forget somebody in the acknowledgments (actions); and 3) The Qualification problem [McCarthy 1977] the problem is that the number of people behind of this PhD research (preconditions for each action) is immense. So, having this in mind, I will take the risk of give my sincere thanks to a very few persons. I would like to express my most sincere gratitude to Helena and Juan Carlos for having trust in me, for being unconditional mentors, for all the guidance and support. You changed my life. Tambi´en agradezco a Patricia, Estefan´ıa y Carolina por hacerme sentir el tan escaso calor y sabor latinoamericano. Es dif´ıcil para mi agradecer con palabras a toda mi familia. A casi todos les tengo un amor incondicional y nombrarlos a todos abarcar´ıa un cap´ıtulo extra. No obstante, quiero agradecer todo el apoyo que me ha dado a mi mam´a, qui´en desde siempre me ha apoyando en cada uno de mis proyectos, te adoro m´a. Y claro, una dedicaci´on especial a mi t´ıo Edmundo Rosero por colaborarme con la portada del libro, con sus obras: “Vendedoras de esperanza” (frente) y “Ancestro” (respaldo). Finally, to Gabrishu for taking me out of my work when more I needed it, bardzo, bardzo cie kocham.

ix

x

Contents

1

1. Introduction 1.1 Research Challenges 1.2 Objectives 1.3 Outline of Thesis

5 7 7 8

2

2. Background 9 2.1 Detecting and Representing Human Activities 9 2.1.1 Hierarchical model of activity complexity 10 2.2 Reasoning about Activities with Incomplete and Inconsistent Information 11 2.3 Tailored Recommendations in Ambient Assisted Living Environments 13

3

3. Contributions

15

4

4. Future Work

19

Part I Publications Paper I

33

Paper II

77

Paper III

89

xi

xii

Glossary Argumentation theory The theory of argumentation is a rich, interdisciplinary area of research spanning across philosophy, communication studies, linguistics, and psychology. Its techniques and results have found a wide range of applications in both theoretical and practical branches of artificial intelligence and computer science. These applications range from specifying semantics for logic programs to natural language text generation to supporting legal reasoning, to decision-support for multi-party human decision-making and conflict resolution [65]. 6 Belief-Desire-Intention model In the study of agent-oriented systems (intelligent agents), a Belief-Desire-Intention architecture views the system as a rational agent having certain mental attitudes of Belief, Desire and Intention, representing, respectively, the information, motivational and deliberative states of the agent [67]. 9 Logic Programming Logic programming is a method for representing declarative knowledge. A logic program consists of rules. A rule has two parts: the head and the body. If the body is nonempty, then we separate it from the head by the symbol ← (“if”): Head ← Body. Major logic programming language families include Prolog, Answer set programming (ASP) and Datalog. When a body of knowledge is expressed as a logic program, logic programming systems can sometimes be used to answer queries on the basis of this knowledge [46]. 6, 11, 12 Ubiquitous Computing Ubiquitous computing has as its goal the non-intrusive availability of computers throughout the physical environment, virtually, if not effectively, invisible to the user [76]. 5

1

2

Acronyms AAF Abstract Argumentation Frameworks. 12 AAL Ambient Assisted Living. 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 19, 20 AI Artificial Intelligence. 5, 9 AmI Ambient Intelligence. 5, 13, 14, 20 ELP Extended Logic Programs. 11 HCI Human-Computer Interaction. 5, 6 KR Knowledge Representation. 13 NAF Negation as Failure. 11 WFS Well-Founded Semantics. 12, 16, 18

3

4

Chapter 1

1. Introduction Ambient Intelligence (AmI) provides a vision of the Information Society where the emphasis is on greater user-friendliness, more efficient service support, userempowerment, and support for human interactions by embedding intelligent behavior in a computerized environment [25]. Within AmI approaches, Ambient Assisted Living (AAL) technologies combine Ubiquitous Computing, intelligent systems and Human-Computer Interaction (HCI) design, placing humans in the center of design, for the purpose of improving every day activities. Different joint European projects1 have pointed out the importance of current AAL and AmI developments for an effective application of assistive technologies for supporting people. In AAL literature, e.g., [1, 5, 25, 76], the key technical requirements which have been commonly identified are the following: intelligent user interfaces, scalable infrastructure, data interoperability, data inconsistency management and the application of a human-centric perspective. Moreover, in both AmI and Artificial Intelligence (AI) literature, different authors [3, 42, 66] have observed that in AmI environments and scenarios, AI methods and techniques can help to accomplish important computational tasks such as: interpreting the environment’s state, representing the information and knowledge associated with the environment, modeling, simulating, and representing entities from the environment, planning decisions or actions and acting on the environment [21]. All these tasks require different sound computational techniques especially developed for accomplishing knowledge representation, reasoning, decision-making and learning. Common-sense reasoning is the type of reasoning that a person often performs when reasoning about what to do, by evaluating the potential results of the different actions (s)he can do [56]. Almost every type of intelligent task (e.g., 1 SOPRANO (Service-oriented Programmable Smart Environments for Older Europeans http://www.soprano-ip.org/), AALIANCE2 (http://aaliance2.eu/), universAAL (UNIVERsal open platform and reference Specification for Ambient Assisted Living http://universaal.org/) and GiraffPlus (http://www.giraffplus.eu/), ISTAG reports (http://cordis.europa.eu/fp7/ict/istag/reports en.html)

5

natural language processing, planning, learning, high-level vision, expert-level reasoning) requires some degree of common-sense reasoning to carry out [21]. The encoding of common-sense knowledge has been recognized as one of the central issues of AI since the inception of the field by authors such as McCarthy [54]. The most important difference between information representations for commonsense reasoning and information representations used in other areas of Computer Science, e.g., relational databases, lies in the expressivity requirement. Most Computer Science representations assume either complete information or information that is partial only along some very limited dimensions. By contrast, common-sense reasoning requires dealing with a wide range of possible types of partial information [21]. Inspired by common-sense reasoning, non-monotonic reasoning captures and represents defeasible inference, i.e., that kind of inference of everyday life in which reasoners draw conclusions tentatively, reserving the right to retract them in the light of further information [4]. In this setting, Argumentation theory has emerged as a formalism for dealing with non-monotonic reasoning. Dung has demonstrated in his seminal paper [26], that many of the major approaches to non-monotonic reasoning in AI and Logic Programming are different forms of argumentation. An artifact, like an AAL system, has common-sense if it automatically deduces a sufficiently wide class of immediate consequences of anything it is told and what it already knows [54]. Hence, in order to capture the individuals perspective of the world, such intelligent machinery must adopt an individual-like or individual-inspired approach or technique for reasoning. Social sciences have successfully developed different frameworks for understanding the human actor and her/his social and physical environment. A broad range of research has shown the underlying processes behind human (commonsense) reasoning [29, 53, 71], the interactions between a person and her/his environment [27, 39, 41, 45, 52, 60] and the social interplay between humans [17, 70, 72]. Furthermore, some studies have also shown that it is not possible to understand human behavior when the world and the social interactions are disregarded [45, 27]. These approaches have had a strong impact in fields like HCI, where the focus for designing artifacts is on the individual and her/his context. Designed artifacts reflect substantial and complex theories of human activity and experience [16, 37], and, at the same time, human-centered design becomes a model of software development [15, 60]. In the remaining part of this chapter, we identify some research challenges for building intelligent artifacts considering the individual’s perspective in AAL. Considering these challenges, we point out the particular research objectives of this licentiate thesis.

6

1.1

Research Challenges

Building an AAL system that is able to successfully perform common-sense reasoning is challenging. The system must be able to capture information from the environment of a person while managing the inherent problems of inconsistency and incompleteness of sensor-based approaches and also to keep track of changes in the individual’s preferences, motives, needs and activity performance. Therefore, the following capabilities are needed by an AAL system: • Detect and represent human activities: Capture knowledge about the individual, the needs and preferences of an individual, in addition to the information about activities that are being performed. • Evaluate the quality of activity performance: Define a quantitative manner to measure the fulfillment (non-fulfillment) degree of an activity and its current status. • Provide tailored services: The AAL system must provide services to an individual, adapted to the individual’s needs, priorities and ability to perform desired activities. In order to achieve these AAL capabilities, the following computational challenges have been identified: • Management of incomplete information, exceptions and inconsistent knowledge bases: Determine a suitable underlying formalism for knowledge representation that is able to manage inconsistency, exceptions and incomplete information. • Define a reasoning model for activity recognition and activity evaluation: Given a knowledge base representing the individual’s interests, the AAL system must provide sound and complete methods of inference, being able to always provide an answer. Therefore, the main research question addressed in this thesis is: How can an AAL system be built, which is based on intelligent artifacts, that performs common-sense reasoning, manages incomplete and inconsistent information about its world and provides tailored services?

1.2

Objectives

Four specific objectives related to the theoretical and technical challenges introduced above have been defined: 1. Build a knowledge representation model based on an argumentation framework for decision-making, which integrates concepts and components from theories of human activity. 7

2. Define a model (a calculus) for evaluating the activity performance, considering an activity as a basic unit of analysis. 3. Implement an argumentation-based reasoner engine for managing incomplete and inconsistent information. 4. Develop a prototype of an AAL system and evaluate the technology with users in a pilot study.

1.3

Outline of Thesis

In the remaining chapters of this thesis, a theoretical background to our results, is introduced in Chapter 3. The achieved contributions to the state of art are presented in Chapter 4. Finally, in Chapter 5, future work addressing extensions of the current work is introduced.

8

Chapter 2

2. Background This chapter provides a theoretical background for the contributions of the thesis, addressing definition and representation of human activities, reasoning about activities with incomplete and inconsistent information and generating tailored recommendations in ambient assisted living environments.

2.1

Detecting and Representing Human Activities

Researchers within the fields of psychology, social and cognitive sciences have developed systematic approaches for studying the human behavior, proposing different methods to describe the interaction of a subject and the world that surrounds her/him, e.g. [63, 77, 44]. In this thesis, the task of representing knowledge about the human being is crucial, both the information about her/his social interaction with the environment, but also regarding needs and motivations. In this regard, a holistic and systemic approach such as Activity Theory seems suitable for the representation of human activities [47, 49]. Furthermore, the structural components that describe a human actor in Activity Theory (motives, goals, needs, actions, etc.) have a significant similarity with one of the most well-known models for modeling a rational agent in AI, the so-called Belief-Desire-Intention model (BDI) [13]. Activity Theory has its threefold historical origins in classical German philosophy (from Kant to Hegel), in the writings of Marx and Engels, and in the Soviet Russian cultural-historical psychology of Vygotsky, Leontiev, and Luria [28]. Nowadays, Activity Theory is transcending its own origins: it has become an international and multidisciplinary research field [27]. Many approaches to AAL use a low-level task, such as sitting, walking, lying down, etc. as the unit of analysis [2, 43, 30]. While this makes it relatively easy to design laboratory experiments, the use of isolated actions in analyzing real-life situations outside a laboratory is less fruitful. 9

In the systemic perspective of Activity Theory, the minimal meaningful context for individual tasks must be included in the basic unit of analysis. This unit is called activity and serves as an important motive for the individual. In our work, this perspective is adopted, striving for identifying meaningful activities at higher levels of complexity.

2.1.1

Hierarchical model of activity complexity

Activity in a narrow sense is a subset of all possible processes related to the interaction of the subject with the world. The subset is defined by its orientation toward a specific motive and what is actually being performed. However, activities are not monolithic. Activity theory defines three types of human actions: activity, consisting of a set of actions, which in turn may consist of actions and operations in a nested structure. An activity constitutes the minimal unit of analysis of purposeful activity and is defined by its motive. Actions have goals and are executed by the actor at a conscious level, in contrast with operations which do not have a goal of their own and which are executed at the lowest level as automated, unconscious processes. The structure is dynamic in that there may be a frequent transformation between the levels, triggered by the demands and prerequisites in the environment or factors in the actor such as lack of knowledge [47]. In Figure 1, the hierarchical structure of an activity is visualized, adapted from [40].

Activity

Action 1

Operation 2.1

Action 2

Operation 2.2

Motive

Action 3

Operation 2.3

Goals

Conditions

Figure 1: Activity Theory hierarchical structure. Analyzing unconsciously performed operations allows us to structure the activity recognition process in a bottom-up manner. By considering framebased models for capturing observations of the world, like those performed by sensor-networks in AAL approaches, we structure an activity based on a set of observations of operations and a model of the human actor, which provides the motives, goals, prioritized activities, habits, etc. In this setting, by adding a topdown view with information about goals and motives of a person to this bottomup activity recognition process, the identification of an activity is performed. The structured hierarchy of an activity is not only important in a recognition process, as further analysis of the activity dynamics can also be performed. Explanation, justification and motivation of meaningful activities constitute different processes of an activity performance analysis following Activity Theory. Reasoning mechanisms are applied in this research to interpret the observations into a set of potential activities which may be taking place. The strongest

10

candidate interpretation is selected and evaluated. This process will be further described in the next section.

2.2

Reasoning about Activities with Incomplete and Inconsistent Information

In this section, a theoretical background to the process of capturing commonsense knowledge from a representation of human activity is introduced, as well as the theoretical basis of an argumentation-based model for non-monotonic reasoning about activities in the presence of incomplete and inconsistent information. Underlying formalisms for representing knowledge structures play an important role, especially in the presence of inconsistent and incomplete information obtained from sensor-based environments. Inconsistency may be present for different reasons, as was highlighted in [10]: 1) contradictory information in the knowledge base (e.g., two AAL sensor-based applications detecting contradictory observations of a human during the performance of an activity); 2) observations conflicting with the normal functioning mode of the system; and 3) discrepancies in the reliability of consistent knowledge bases (e.g., different sensor-based services providing the location of a person, with different levels of accuracy). A large number of “non-monotonic” reasoning approaches have been developed for capturing common-sense knowledge [55, 57, 59, 68, 51, 38, 61]. Among them, Extended Logic Programs (ELP) [31] captures incomplete information as well as exceptions, using classical negation and Negation as Failure (NAF). NAF was introduced as a 1-place connective (not) when proportional symbols are used, e.g. ¬q ← not q, which should be understood as “q is false if there is no evidence that p is true”. On programs with NAF, the consequence operator is not monotonic, changing the evaluation result as more information is added to the program. ELP are used for managing incomplete knowledge in a knowledge base by extending logic programs with a second type of negation ¬ called “classical”, “strong”, or “explicit” by different authors who associate different meanings to it, in addition to NAF [8]. By applying ELP as the underlying formalism for knowledge representation in our approach, we can take advantage of its model-based semantics for capturing inconsistent and incomplete information. For the purpose of reasoning with the inconsistent and incomplete information represented using ELP, we have two alternative semantics to use, which have different qualities: • Answer Set Semantics [31], an extension of Stable model semantics: Logic Programming with answer set models constitutes one of the most efficient and widely used approaches of non-monotonic reasoning for reasoning with 11

incomplete and uncertain information [7]. An answer set is represented by a collection of ground literals. Informally, a ground program can be viewed as a specification for the sets of beliefs which could be held by a rational reasoner associated with such a program. • a version of the Well-Founded Semantics (WFS) [73]: WFS performs skeptical reasoning1 , and data complexity of logic programs under WFS is polynomial and is defined for all general logic programs [73, 8]. These characteristics allow us to evaluate programs in a feasible manner. Moreover, it is worth mentioning that WFS is a skeptical approximation of stable models. WFS was selected for our approach due to its non-monotonic reasoning properties2 and its qualities relating to its computational complexity, among other properties. Dung made an important contribution to the argumentation literature in [26], showing that Logic Programming (LP) can be expressed as a form of argumentation inference, providing a meta-theory of such systems. The analysis of non-monotonic reasoning based on argumentation shares three major points with common-sense reasoning [21]: 1. Representation: The development of knowledge structures that can express facts in the domain. 2. Domain theory: The characterization of the fundamental properties of the domain and the rules that govern it. 3. Inference techniques: The construction of algorithms or heuristics that can be used to automate useful types of reasoning. Dung defined a meta-interpreter [26], which follows this three-point vision of a common-sense reasoner, Figure 2. This meta-interpreter has been the high level model for different Abstract Argumentation Frameworks (AAF) ([26, 12], among others), providing a theoretical basis for exploring issues of defeasible reasoning 3 . In our approach, an argument-based reasoner based on Dung’s meta interpreter was implemented. We extended Dung’s model by adding a conclusion inference, which takes information about the individual into consideration, and by defining and implementing the notion of global selection (Paper II). This “three-step” process starts with a knowledge base with common-sense knowledge from a human’s perspective and ends with a conclusion or a sound set of conclusions about an observed activity performed by a human (Figure 2). Our 1 Well-Founded Semantics is an inherently skeptical semantics that abstains from drawing conclusions whenever there is a potential conflict in the knowledge base representation [14]. 2 Dix showed that WFS is a well-behaved semantics [23]. 3 Reasoning is defeasible, in the sense that the premises accept a conclusion, but when additional information is added, that conclusion may no longer be justified [64]

12

Knowledge Base

Step 1

Step 2

Step 3

Argument Building

Argument Acceptance Analysis

Conclusion Inference

Argument Framework

Argument Extensions

attacks set attacks set attacks set

acceptable attacksset set attacks arguments

Argument-based conclusion

set

General Argument Meta Interpreter by Dung

Figure 2: Argument-based reasoner based on the Dung’s meta-interpreter. implementation follows this meta-architecture, providing sound proofs of goalbased human activities, using a decision-making framework (Step 1 and Step 2) and a human-centric explanation of an activity (Step 3) based on activitytheoretical models of human activity (Paper I).

2.3

Tailored Recommendations in Ambient Assisted Living Environments

In AAL and AmI literature, a large number of approaches use information about the human and her/his environment to personalize the provision of services. During the process of generating a tailored recommendation in AmI systems, some approaches fuse information in a Knowledge Representation (KR) phase from different sources, as is shown in Figure 3 [19, 74].

Figure 3: Information flow in ambient intelligence, adapted from [18]. In the KR phase, there are important requirements to consider. Models of human behavior and cognition should be represented using KR techniques. Topological and physical aspects of the environment should be easily encoded. Several of the approaches use stereotypical patterns of human reasoning, called schemes in argumentation literature [33, 75]. These reasoning scheme structures 13

are also used in argumentation approaches for persuasion [58, 9, 62, 32] and for building arguments or locutions, based on information about the individual. Another model for using information about an individual in AmI literature is based on research in the field of Recommender Systems. Recommender Systems typically implements e-services, which use personal information to advertise and recommend objects. By using information of a person as a parameter for selection after a filtering process, some AmI approaches are close to recommender systems [22, 69, 43, 30, 2]. In Paper III, two processes are introduced for choosing the “best option” in an argument-based system: the local selection and global selection, which build a hybrid approach for tailoring a recommendation, modelling the knowledge base with information of a user, and also using data obtained from observations of human activity for the global selection. In the local selection process, the fragmented information about the human activity, obtained in an AAL environment is modeled into a knowledge base. The global selection process interprets the content of the knowledge base by fusing this information with information modelled at a higher level of activity obtained in encounters between the individual and therapists or in dialogues between the individual and a dialogue system.

14

Chapter 3

3. Contributions In this chapter, we describe contributions with respect to the main goal of this thesis: building intelligent machinery for supporting human activities in AAL environments.

Paper I: Arguing through the Well-founded Semantics: an Argumentation Engine In paper I, we introduced different technical contributions to the state of art, in addition to an implementation of an argument-based reasoner (Objective 3 in Section 1.2). The paper focuses on the development of a framework for building arguments meeting the following requirements: 1. management of inconsistency and incompleteness in the knowledge base, 2. following consistency principles, 3. efficient (conclusions should be computed in polynomial time), 4. affordable prototyping. In this setting, we developed a method for building consistent arguments, which uses as input any extended logic program, i.e., a program capturing incomplete information and exceptions, using classical negation and negation as failure, introduced in Section 2.2. The resulting arguments have the following novel characteristics: 1. the information for supporting a conclusion in our argument structure can handle not-stratified programs 1 , endowing our approach with a tractable management of inconsistent knowledge bases. 1 Stratified

programs is a class of logic programs in which most of the well-accepted logic programming semantics coincide

15

2. arguments follow a set of rationality postulates used to judge the quality of a rule-based argumentation system. 3. the computational complexity of the system is maintained in the polynomial class. 4. the knowledge representation captures information from the individual as well as from the AAL environment, achieving in this way Objective 1. By taking advantage of the computational properties of the Well-Founded Semantics, we developed an open-source implementation2 , corresponding to the Reasoner Agent in Figure 4 (Objective 3, 4). Furthermore, Well-Founded Semantics follows the relevance principle3 [23], allowing us a special treatment of a knowledge base by evaluating only relevant rules, and, in this way, to identify valid sets of rules suitable for creating arguments.

Paper II: Reasoning about Human Activities: an Argumentative Approach In Paper II, a novel approach for evaluating activity performance, based on an integration of argumentation theory and models from Activity Theory is presented. Moreover, we introduce a calculus for evaluating the performance of human activities, in order to estimate quantitatively the achievement of composing goals of an activity (Objective 2). This approach, as far as we know, is the first proposed argumentation-theoretical method for evaluating the performance of an activity that considers a human-centric perspective. We define the concept of hypothetic fragments of activities, following the ideas of Activity Theory, which suggests that an activity is motivated by needs. Furthermore, the method for analyzing human activities is also novel, introducing a two-step procedure for making the selection of the best alternatives for obtaining an activity explanation, consisting of the following: Step 1) local selection, extending the Dung’s meta-interpreter for selecting hypothetic fragments of activities; and Step 2) global selection, defining rules for determining which sets of hypothetical fragments of activities could form evidence about the fulfillment or non-fulfillment of some particular activity. In this paper, the behavior of generic argumentation semantics and Dung’s argumentation semantics w.r.t. activity recognition is studied. This calculus is intended to be the underlying approach to evaluating the performance in AAL systems. 2 The

“Argument Engine” is available in: http://www8.cs.umu.se/∼esteban/wfsArgEngine (source and manuals) 3 Relevance is a property that WFS satisfies [23] and states, informally speaking, that: it is perfectly reasonable that the truth-value of an atom, with respect to any semantics, only depends on the subprograms formed from the relevant clauses with respect to that specific atom [24]

16

ALI System (Paper III) ALI Mobile Application on{X} HTTP

Data Sensing Notification Delivery

Argument Engine (Paper I)

Activity Performance Measurement (Paper II) HTTP

ACKTUS ACKTUS - Data Capture App

Activity and Goals Definition

Figure 4: ALI Architecture integrated with ACKTUS repositories.

Paper III: ALI, an Ambient Assisted Living System for Supporting Behavior Change In Paper III, we present a prototype of an AAL system ALI: an Assisted LIving system, integrating the argument engine introduced in Paper I. Our approach is exemplified in a case scenario of an individual living in an AAL environment. Methods for tracking and monitoring a person using a sensor-based mobile application ALI-MOB are presented (Figure 4). ALI generates recommendations using information from the observations obtained from the person’s environment and from goal-oriented activities, which were selected and prioritized by the person. ALI is integrated with ACKTUS4 knowledge repositories and applications 4 ACKTUS (Activity-Centered Knowledge and interaction modeling Tailored to USers) is an evolving semantic web application that is designed to allow domain experts who are typically not familiar with knowledge engineering to author and model the knowledge content of and design the interaction with knowledge-based applications [50]. The knowledge-based support applications described and used in our case study were developed by domain professionals using ACKTUS [48, 49].

17

[48, 49], which: 1) maintains the user information; 2) interacts with a healthcare team; and 3) provides tools for assessment and user evaluation. In this stage of the ALI development, the integration with ACKTUS is limited to an exchange of information, with no integration of the reasoning models. The implementation of ALI had two main components: 1) a centralized system, containing mainly the reasoner and recommendation agents (based on the WFS argument engine); and 2) the mobile sensor-based application (Figure 4). Two more papers were published [34, 35] introducing different application context possibilities of ALI. Furthermore, a pilot evaluation study, which was conducted using ALI to monitor young adolescents and to provide recommendations for achieving a healthier lifestyle, provided results which have been fed into further developments of ALI. The results are submitted for publication and builds a base for an extended user study to be conducted as future work, addressing Objective 4 (Section 1.2).

18

Chapter 4

4. Future Work In this chapter, we delineate the future extensions of our current work and describe some ongoing work connected with the contributions introduced in previous chapters. Considering the proposed research question and the achieved objectives so far, we will focus on the following research topics as future work: • Extensions of the ALI platform: In Paper III and in [34, 35], the importance of integration with more knowledge repositories was evident. A lesson learned from collaborative projects with physiotherapists is that integrating and adding more information for evaluating the activity performance is important. Moreover, models of reasoning integrating expert opinions residing in the ACKTUS repositories improve the quality of the activity evaluation. Therefore, the following will be subjected to future work: – Full integration with ACKTUS knowledge repositories: providing in this way a full coverage of the three common-sense reasoning topics (representation, domain theory, inference techniques), focused on AAL environments. – An implementation of the management of goal-preferences via a possibilistic decision-making framework and the person’s preferences obtained from ACKTUS. • Evaluation of the ALI platform: a pilot study has been conducted, aimed at evaluating ALI as a health activity recommender. The results have been fed into the technical solution and ALI will be further evaluated with a larger use group over a longer period of time. – A long term evaluation study using mobile and wearable devices in order to capture the environment and activities of the person. An initial pilot study will be focused on assessing balance in older adults in different activities.

19

– Short term pilot studies for testing techniques for automatic configuration of collaborative agents in an AAL environment which utilize ACKTUS. • Stream reasoning in the Activity Argumentation Framework: In different contexts like AmI and AAL, common-sense reasoning requires the representation of time. Temporal Argumentation Frameworks and argumentation using temporal knowledge have been introduced in the argumentation literature [20, 6, 11, 20]. We want to extend those approaches integrating our strategy for evaluating activity performance, which also handles the management of inconsistency and incomplete knowledge that we developed in Paper I. We are working on an agent-based middle-ware platform for providing real time services in an AAL context (see Figure 5).

Figure 5: Developed software platforms implementing the research presented in this thesis. We have developed and tested approaches for tracking and monitoring human activities in different contexts, for instance, a mobile application for monitoring exposure to vibration in a work-environment. This application is intended to be integrated with ACKTUS in order to detect and recommend safety practices in the mining and construction industries. The principle for tracking, monitoring vibration levels and recommending is similar to the one used in Paper III and in [34, 35]. 20

In collaboration with physiotherapists, we are currently developing a monitoring approach for detecting and evaluating balance in the elderly. The mobile sensor-based application tracks the movement of a person during a balance test organized by physiotherapists, using a version of the Short Physical Performance Battery Score test [36]. The idea behind these two projects is to test different approaches for improving the sensor-based technologies, as well as the enhancement of methods for dealing with the inherent problem these technologies need to be able to handle, namely managing inconsistent and incomplete information. In Figure 5, we summarize the software architectures and prototypes which implements the research results of this thesis.

21

22

Bibliography [1] Gregory D Abowd and James PG Sterbenz. Final report on the interagency workshop on research issues for smart environments. Personal Communications, IEEE, 7(5):36–40, 2000. [2] M Amoretti, F Wientapper, and F Furfari. Sensor data fusion for activity monitoring in ambient assisted living environments. In ICST 2010, pages 206–221. Springer, Pisa, Italy, 2010. [3] Aliaksei Andrushevich, Stephan Tomek, and Alexander Klapproth. The autonomic computing paradigm in adaptive building/ambient intelligence systems. In Ambient Intelligence, pages 98–104. Springer, 2011. [4] G Aldo Antonelli. Non-monotonic logic. Stanford Encyclopedia of Philosophy, 2006. [5] Pablo Oliveira Antonino, Daniel Schneider, Cristian Hofmann, and Elisa Yumi Nakagawa. Evaluation of aal platforms according to architecture-based quality attributes. In Ambient Intelligence, pages 264– 274. Springer, 2011. [6] Juan C Augusto and Guillermo R Simari. A temporal argumentative system. AI Communications, 12(4):237–257, 1999. [7] Chitta Baral. Knowledge Representation, Reasoning and Declarative Problem Solving. Cambridge University press, 2003. [8] Chitta Baral and Michael Gelfond. Logic programming and knowledge representation. The Journal of Logic Programming, 19:73–148, 1994. [9] Trevor JM Bench-Capon. Persuasion in practical argument using valuebased argumentation frameworks. Journal of Logic and Computation, 13(3):429–448, 2003. [10] Salem Benferhat, Didier Dubois, and Henri Prade. Some syntactic approaches to the handling of inconsistent knowledge bases: A comparative study part 1: The flat case. Studia Logica, 58(1):17–45, 1997.

23

[11] Philippe Besnard et al. Argumentation using temporal knowledge. Computational Models of Argument: Proceedings of COMMA 2008, 172:204, 2008. [12] Andrei Bondarenko, Phan Minh Dung, Robert A Kowalski, and Francesca Toni. An abstract, argumentation-theoretic approach to default reasoning. Artificial Intelligence, 93(1):63–101, 1997. [13] M. Bratman. Intention, plans, and practical reason. Harvard University Press, 1987. [14] Gerhard Brewka. Well-founded semantics for extended logic programs with dynamic preferences. Journal of Artificial Intelligence Research, 4:19–36, 1996. [15] John M Carroll. Human–computer interaction: Psychology as a science of design. International Journal of Human-Computer Studies, 46(4):501–522, 1997. [16] John M Carroll and Robert L Campbell. Artifacts as psychological theories: The case of human-computer interaction. Behaviour & Information Technology, 8(4):247–256, 1989. [17] Tanya L Chartrand and John A Bargh. The chameleon effect: The perception–behavior link and social interaction. Journal of Personality and Social Psychology, 76(6):893, 1999. [18] Nak-Young Chong, Fulvio Mastrogiovanni, and IGI Global. Handbook of Research on Ambient Intelligence and Smart Environments: Trends and Perspective. Information Science Reference, 2011. [19] Marcello Cirillo, Federico Pecora, and Alessandro Saffiotti. Proactive assistance in ecologies of physically embedded intelligent systems: a constraintbased approach. Handbook of Research on Ambient Intelligence and Smart Environments: Trends and Perspective, 370:534–557, 2011. [20] Maria Laura Cobo, Diego C Mart´ınez, and Guillermo Ricardo Simari. On admissibility in timed abstract argumentation frameworks. In ECAI, volume 215, pages 1007–1008, 2010. [21] Ernest Davis. Pouring liquids: A study in commonsense physical reasoning. Artificial Intelligence, 172(12):1540–1578, 2008. [22] H´ector Delgado, Aitor Rodriguez-Alsina, Antoni Gurgu´ı, Enric Mart´ı, Javier Serrano, and Jordi Carrabina. Enhancing accessibility through speech technologies on aal telemedicine services for itv. In Ambient Intelligence, pages 300–308. Springer, 2011. [23] J¨ urgen Dix. A classification theory of semantics of normal logic programs: II. weak properties. Fundamenta Informaticae, 22(3):257–288, 1995. 24

[24] J¨ urgen Dix and Martin M¨ uller. Partial evaluation and relevance for approximations of stable semantics. In ISMIS, pages 511–520, 1994. [25] Ken Ducatel, Marc Bogdanowicz, Fabiana Scapolo, Jos Leijten, and JeanClaude Burgelman. Scenarios for ambient intelligence in 2010. Office for official publications of the European Communities, 2001. [26] Phan Minh Dung. On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games. Artificial Intelligence, 77(2):321–357, September 1995. [27] Y. Engestr¨om, R. Miettinen, and R.L. Punam¨ aki. Activity theory and individual and social transformation. In Perspectives on Activity Theory, Learning in Doing: Social, Cognitive and Computational Perspectives, pages 19–38. Cambridge University Press, 1999. [28] Yrj¨o Engestr¨om. Developmental work research: Expanding activity theory in practice, volume 12. Lehmanns Media, 2005. [29] Jonathan St BT Evans, Stephen E Newstead, and Ruth MJ Byrne. Human Reasoning: The Psychology of Deduction. Psychology Press, 1993. [30] Chrysi Filippaki, Grigoris Antoniou, and Ioannis Tsamardinos. Using Constraint Optimizationfor Conflict Resolution and Detail Control in Activity Recognition. In AmI 2011, pages 51 – 60. Springer, 2011. [31] Michael Gelfond and Vladimir Lifschitz. Classical negation in logic programs and disjunctive databases. New Generation Computing, 9(3-4):365– 385, 1991. [32] Floriana Grasso, Alison Cawsey, and Ray Jones. Dialectical argumentation to solve conflicts in advice giving: a case study in the promotion of healthy nutrition. International Journal of Human-Computer Studies, 53(6):1077– 1115, 2000. [33] W. Grennan. Informal Logic: Issues and Techniques. McGill-Queen’s Press - MQUP, 1997. [34] Esteban Guerrero, Helena Lindgren, and Juan Carlos Nieves. Ali, an assisted living system based on a human-centric argument-based decision making framework. In 13th Workshop on Computational Models of Natural Arguments (CMNA 2013), pages 46–51, 2013. [35] Esteban Guerrero, Juan Carlos Nieves, and Helena Lindgren. Ali: An assisted living system for persons with mild cognitive impairment. In Computer-Based Medical Systems (CBMS), 2013 IEEE 26th International Symposium on, pages 526–527. IEEE, 2013.

25

[36] Jack M Guralnik, Eleanor M Simonsick, Luigi Ferrucci, Robert J Glynn, Lisa F Berkman, Dan G Blazer, Paul A Scherr, and Robert B Wallace. A short physical performance battery assessing lower extremity function: association with self-reported disability and prediction of mortality and nursing home admission. Journal of Gerontology, 49(2):M85–M94, 1994. [37] Langdon inner. Do artifacts have politics? In Technology, Organizations and Innovation: Theories, concepts and paradigms, pages 121–136. Taylor & Francis, 1980. [38] Antonis C. Kakas, Robert A. Kowalski, and Francesca Toni. Abductive logic programming. Journal of Logic and Computation, 2(6):719–770, 1992. [39] Victor Kaptelinin and Marc Hassenzahl. Activity theory. The Encyclopedia of Human-Computer Interaction, 2nd Ed., 2013. [40] Victor Kaptelinin and Bonnie Nardi. Acting with Technology: Activity Theory and Interaction Design. Acting with Technology. MIT Press, 2006. [41] Victor Kaptelinin and Bonnie A Nardi. Activity theory: Basic concepts and applications. In CHI’97 Extended Abstracts on Human Factors in Computing Systems, pages 158–159. ACM, 1997. [42] Jeffrey O Kephart and David M Chess. The vision of autonomic computing. Computer, 36(1):41–50, 2003. [43] Eunju Kim and Sumi Helal. Revisiting human activity frameworks. In ICST Conf., S-Cube 2010, pages 219 – 234, Miami, FL, USA, 2010. Springer. [44] Aleksi Nikolaevich Leontiev. Activity, consciousness, and personality. Prentice-Hall, Englewood Cliffs, N.J., 1978. [45] Aleksei Nikolaevich Leontyev. Activity and consciousness. Moscow: Personality, 1974. [46] Vladimir Lifschitz. Foundations of logic programming. Principles of Knowledge Representation, 3:69–127, 1996. [47] Helena Lindgren. Activity-theoretical model as a tool for clinical decision-support development. In 15th International Conferenceon KnowledgeEngineeringand KnowledgeManagement, Poster and demo proceedings in EKAW, volume 215, pages 23–25, 2006. [48] Helena Lindgren, Lillemor Lundin-Olsson, Petra Pohl, and Marlene Sandlund. End users transforming experiences into formal information and process models for personalised health interventions. Studies in Health Technology and Informatics, 205:378–82, 2014. [49] Helena Lindgren and Ingeborg Nilsson. Towards user-authored agent dialogues for assessment in personalised ambient assisted living. International Journal of Web Engineering and Technology, page in press, 2013. 26

[50] Helena Lindgren, Patrik J. Winnberg, and Peter Winnberg. Domain experts tailoring interaction to users - an evaluation study. In INTERACT (3), pages 644–661, 2011. [51] Jorge Lobo, Jack Minker, and Arcot Rajasekar. Foundations of disjunctive logic programming. The MIT Press, 1992. [52] Lorenzo Magnani and Emanuele Bardone. Ambient Intelligence as Cognitive Niche Enrichment. In Handbook of Research on Ambient Intelligence and Smart Environments, chapter 1, pages 1–17. IGI Global, 2011. [53] Norman RF Maier. An aspect of human reasoning. British Journal of Psychology. General Section, 24(2):144–155, 1933. [54] John McCarthy. Programs with common sense. Defense Technical Information Center, 1963. [55] John McCarthy. Circumscription: a form of non-monotonic reasoning. Artificial Intelligence, 13(1):27–39, 1980. [56] John McCarthy and Patrick Hayes. Some philosophical problems from the standpoint of artificial intelligence. Stanford University USA, 1968. [57] Drew McDermott. Nonmonotonic logic II: Nonmonotonic modal theories. Journal of the ACM (JACM), 29(1):33–57, 1982. [58] Maria Miceli, Fiorella de Rosis, and Isabella Poggi. Emotional and nonemotional persuasion. Applied Artificial Intelligence, 20(10):849–879, 2006. [59] Robert C Moore. Semantical considerations on nonmonotonic logic. Artificial Intelligence, 25(1):75–94, 1985. [60] Bonnie A Nardi. Context and Consciousness: Activity Theory and HumanComputer Interaction. Mit Press, 1996. [61] Donald Nute. Defeasible logic. In Web Knowledge Management and Decision Support, pages 151–169. Springer, 2003. [62] Chaim Perelman. The new rhetoric: A treatise on argumentation. In Yehoshua Bar-Hillel, editor, Pragmatics of Natural Languages, volume 41 of Synthese Library, pages 145–149. Springer Netherlands, 1971. [63] Jean Piaget. The origins of intelligence in children. Journal of Consulting Psychology, 17(6):467, 1953. [64] John L Pollock. Defeasible reasoning. Cognitive Science, 11(4):481–518, 1987. [65] Iyad Rahwan. Guest editorial: Argumentation in multi-agent systems. Autonomous Agents and Multi-Agent Systems, 11(2):115–125, 2005.

27

[66] Carlos Ramos, Juan Carlos Augusto, and Daniel Shapiro. Ambient intelligence, the next step for artificial intelligence. Intelligent Systems, IEEE, 23(2):15–18, 2008. [67] Anand S Rao, Michael P Georgeff, et al. Bdi agents: From theory to practice. In ICMAS, volume 95, pages 312–319, 1995. [68] Raymond Reiter. A logic for default reasoning. Artificial Intelligence, 13(1):81–132, 1980. [69] Ryo Sakai, Sarah Van Peteghem, Leoni van de Sande, Peter Banach, and Maurits Kaptein. Personalized persuasion in ambient intelligence: The apstairs system. In Ambient Intelligence, pages 205–209. Springer, 2011. [70] Mark Snyder and William B Swann Jr. Behavioral confirmation in social interaction: From social perception to social reality. Journal of Experimental Social Psychology, 14(2):148–162, 1978. [71] Keith Stenning and Michiel Van Lambalgen. Human Reasoning and Cognitive Science. MIT Press, 2008. [72] Harry C Triandis. The self and social behavior in differing cultural contexts. Psychological Review, 96(3):506, 1989. [73] Allen Van Gelder, Kenneth A. Ross, and John S. Schlipf. The well-founded semantics for general logic programs. Journal of the ACM, 38(3):619–649, July 1991. [74] Martijn H Vastenburg and Natalia Andrea Romero Herrera. Experience tags: enriching sensor data in an awareness display for family caregivers. In Ambient Intelligence, pages 285–289. Springer, 2011. [75] Douglas Walton, Christopher Reed, and Fabrizio Macagno. Argumentation schemes. Cambridge University Press, 2008. [76] Mark Weiser. The computer for the 21st century. Scientific American, 265(3):94–104, 1991. [77] K. Wolfgang. The Mentality of Apes. Routledge, 2013.

28

Part I

Publications

29

30

I

Arguing through the Well-founded Semantics: an Argumentation Engine Esteban Guerrero, Juan Carlos Nieves, Helena Lindgren Department of Computing Science Ume˚ a University SE-901 87, Ume˚ a, Sweden

Abstract In this paper, we introduce an argumentation approach which takes an extended logic program as input and gives a set of arguments with their respective disagreements as output. We establish the concept of argument under the WellFounded Semantics evaluation, allowing us to identify arguments with stratified programs as support, even when the input for the argument engine is a nonstratified program. We propose a set of rationality postulates for argumentbased systems under extended logic programs, which are based on a definition of closure for a set of clauses that consider the well-known Gelfond and Lifschitz reduction. We establish the conditions in which our approach satisfies these principles. In addition to introducing the algorithms for implementing our approach, we present a standalone argumentation-tool based on the XSB system. Keywords: argumentation, logic programming, well-founded semantics, argumentation tools.

1. Introduction Commonsense knowledge representation is a complex task, particularly under traditional logic programming languages. In argumentation literature, there Email address: (esteban,jcnieves,helena)@cs.umu.se (Esteban Guerrero, Juan Carlos Nieves, Helena Lindgren)

Preprint submitted to Elsevier

September 19, 2014

are a number of approaches based on logic programming (e.g [1, 2, 3]), extending 5

logic programming (e.g. [4, 5, 6]), or based upon classical logic (e.g. [7, 8]) (an extended survey was introduced in [9]). Extended logic programs (ELP) [10], use both strong negation ¬ and negation-as-failure not, allowing us to represent commonsense knowledge by logic programs, dealing directly with incomplete information, as its (axiom system) syntax and semantics is more complete and

10

computationally superior, instead of predicate calculus and circumscription [11]. Two major semantics for ELP have been defined: (1) answer set semantics [10], an extension of Stable model semantics, and (2) a version of the Well-Founded Semantics (WFS) [12]. The second approach can be viewed as an efficient approximation of the first one [13].

15

In the argumentation reasoning context, Dung in his seminal work [14] as well as other [5, 15, 16], define Abstract Argumentation Frameworks providing basis and analytic tools for non-monotonic reasoning, disregarding the structure or nature of the arguments themselves. At the same time, different approaches [17, 18, 19, 20, 21] have instantiated these frameworks in argument-based sys-

20

tems (ABS) (building arguments, applying a given semantics and identifying accepted conclusions). In this setting, the underlying formalisms for knowledge representation play an important role in building arguments. Despite the fact that a number of “non-monotonic” logics [22, 23, 24, 25] have been developed for capturing commonsense knowledge, there is a limited availability of ABS

25

formalisms and their implementations showing a practical use in scenarios with incomplete information. Moreover, some of the ABS mentioned above fail to satisfy logical closure and consistency, which are important properties in any logical reasoning system, as has been pointed out by various authors [26, 27]. Indeed, Caminada formalizes a set of rationality postulates focused on an ABS

30

family [28, 29]. Furthermore, reduced research (to our knowledge only [27]) has been done to study closure and consistency under other well-known ABS under logic programming approaches. Against this background, we introduce an argumentation approach for building arguments based on the WFS. Our approach takes as a knowledge base 2

35

captured by an extended logic program input, obtaining as a result a set of consistent arguments, where an argument entails a relationship support-conclusion. In addition, we introduce a set of rationality postulates for ABS under extended logic programs. These postulates satisfy the existent lack of rationality principles in terms of logic programming approaches.

40

Essentially, our argumentation engine differs from previous approaches in: (1) our language for representing knowledge is based on extended logic programs [10], allowing us to deal in a natural and convenient way with incomplete information; (2) we take advantage of remarkable properties of a skeptical nonmonotonic reasoning semantics: the WFS [12]; (3) the proposed rationality

45

postulates are based on the well-known Gelfond and Lifschitz reduction [30], which is the core for defining the Stable model semantics. WFS captures the common sense notion of negation as failure, leading to a natural treatment of exceptions. The relevance property which is satisfied by WFS, allows us to identify only those clauses relevant to building a stratified support inferring a

50

conclusion, avoiding problems associated with the so-called conflict propagation [27] or contamination [31, 32]. Furthermore, because WFS is polynomial time computable, we developed an implementation which, to our knowledge, is the first argumentation engine using this approach. In summary, the following technical contributions are made:

55

ˆ a notion for an argument is proposed, with a stratified program as support,

using as input any extended logic program (stratified or not). ˆ a set of rationality postulates for argument-based systems based on ex-

tended logic programs. ˆ a prototype tool which may be downloaded and tested. 60

The rest of the paper is structured as follows. In Section 2, we recall the theoretical background used throughout the paper. We introduce the foundational definition of an argument under WFS in Section 3. Rationality postulates under ELP are explored in Section 4. Our argumentation tool and its architecture are 3

introduced in Section 5. We present a discussion about the related work in the 65

state-of-the-art in Section 6. In Section 7 a summary with our contributions and future work is presented. A definition of the Well-Founded Semantics in terms of rewriting systems is introduced in Appendix A. In Appendix B, proofs are presented for considered propositions. The Appendix C section introduces a set of algorithms for building arguments. Appendix D presents a full example for

70

building arguments using our approach and Appendix E describes a practical use case for our argument approach.

2. Background In this section some background about logic programs is introduced. We assume that the reader is familiar with basic terms about Logic Programming 75

with negation as failure. A basic introduction about this topic can be found in [33] 2.1. Logic Programs Let us introduce the language of a propositional logic, which is constituted by proportional symbols: p0 , p1 , . . . ; connectives: ∧, ←, ¬, not, >; and auxiliary symbols: ( , ), in which ∨, ∧, ← are 2-place connectives, ¬, not are 1-place connectives and > is a 0-place connective. The propositional symbols, > and the propositional symbols of the form ¬pi (i ≥ 0) stand for the indecomposable propositions, which we call atoms 1 , or atomic propositions. The negation symbol

¬ is regarded as the so called strong negation in the Answer Set Programming literature, and the negation symbol not as negation as failure. A literal is an atom, a (called a positive literal), or the negation of an atom not a (called a negative literal). A (propositional) extended normal clause, C, is denoted: a ← b1 ∧ · · · ∧ bj ∧ not bj+1 ∧ · · · ∧ not bj+n 1 The

(1)

atoms of the form ¬a are also called extended atoms in the literature. In order to

simplify the presentation, we call them atoms as well.

4

in which j + n ≥ 0, a is an atom and each bi (1 ≤ i ≤ j + n) is an atom. When j + n = 0 the clause is an abbreviation of a ← > (a fact), such that 80

> is the propositional atom that always evaluate to true. In a slight abuse of notation, we sometimes write the clause (1) as a a ← B + ∧ not B − , where

B + := {b1 , . . . , bj } and B − := {bj+1 , . . . , bj+n }. An extended logic program P is a finite set of extended normal clauses. When n = 0, the clause is called an extended definite clause. An extended definite logic program is a finite set 85

of extended definite clauses (e.g., Figure 1 left). By LP , we denote the set of atoms which appear in P. Let us introduce a concept which establishes a notion of dependency between atoms from LP . This concept, is defined as follows: Definition 1. [34] Given an extended logic program P , a relationship: a de-

90

pends immediately on b, if and only if, b appears in the body of a clause in P , such that a appears in its head. The set of dependencies of an atom x, denoted by dependencies of(x), corresponds to the set {a | x depends on a}. The twoplace relationship depends on is the transitive closure of depends immediately on.

95

The concept of dependency as described in Definition 1 is essential for analyzing connectivity between clause literals. Problems emerge when positive and negative literals are connected generating cycles, like in program P introduced in Figure 1, where c and d are “deadlocked”, each waiting for the other to either succeed or fail. Cycles through negation has been analyzed in logic

100

programming literature in various works [35, 36, 37].

c

d

a

n

Figure 1: An extended definite logic program

5

An extended logic program in which negative interactions among literals are acyclic, will be called stratified and is defined as follows: Definition 2. [34] An extended logic program P is called stratified, if it is decomposable as P = P1 ∪ · · · ∪ Pn , such that the following holds (for i = 105

1, . . . , n): 1. If a literal x occurs positively in a clause in Pi , then all clauses containing S x in their heads are contained in j≤i Pj .

2. If a literal x occurs negatively in a clause in Pi , then all clauses containing S x in their heads are contained in j ∈ P } and P f alse := {p| p ∈ LP \HEAD(P )}. SEM(P ) is also called a model of P.

In order to present a characterization of the well-funded semantics in terms of rewriting systems, we define some basic transformation rules for normal logic 715

programs. Definition 13 (Basic Transformation Rules). [54] A transformation rule is a binary relation on ProgL . The following transformation rules are called basic. Let a program P ∈ ProgL be given. RED+ : This transformation can be applied to P if there is an atom a which

720

does not occur in HEAD(P). RED+ transforms P to the program where all occurrences of not a are removed. RED− : This transformation can be applied to P , if there is a rule a ← > ∈ P .

RED− transforms P to the program where all clauses that contain not a in their bodies are deleted.

725

Success: Suppose that P includes a fact a ← > and a clause q ← body such that a ∈ body. Then we replace the clause q ← body by q ← body \ {a}.

30

Failure: Suppose that P contains a clause q ← body such that a ∈ body and a∈ / HEAD(P ). Then we erase the given clause. Loop: We say that P2 results from P1 by LoopA if, by definition, there is a set 730

A of atoms such that: 1. for each rule a ← body ∈ P1 , if a ∈ A, then body ∩ A 6= ∅, 2. P2 := {a ← body ∈ P1 |body ∩ A = ∅}, 3. P1 6= P2 . Let CS0 be the rewriting system such that it contains the transformation

735

rules: RED+ , RED− , Success, F ailure, and Loop. We denote the uniquely determined normal form of a program P with respect to the system CS0 by norm CS0 (P ). Every system CS0 induces a semantics SEMCS0 as follows: SEMCS0 (P ) := SEM(norm CS0 (P )).

740

In order to illustrate the basic transformation rules, let us consider the following example. Example 5. Let P be the following normal program: d(b) ← not d(a).

d(c) ← not d(b).

d(c) ← d(a).

Now, let us apply CS0 to P . Since d(a) ∈ / HEAD(P ), then, we can apply

745

RED+ to P . Thus we get: d(b) ← >.

d(c) ← not d(b).

d(c) ← d(a).

Notice that we can now apply RED− to the new program, thus we get: d(b) ← >. d(c) ← d(a). Finally, we can apply Failure to the new program, thus we get: d(b) ← >. 750

This last program is called the normal form of P w.r.t. CS0 , because none of the transformation rules from CS0 can be applied. WFS was introduced in [55] and was characterized in terms of rewriting systems in [56]. This characterization is defined as follows: 31

Lemma Appendix A.1. [56] CS0 is a confluent rewriting system. It induces 755

a 3-valued semantics that it is the Well-founded Semantics. For instance, in Example 5 we saw that the normal form of P is: d(b) ← >. Hence W F S(P ) = h{d(b)}, {d(a), d(c)}i Appendix B. Proofs

Proposition 1 Let P be an extended logic program. If hS, gi ∈ ArgP , then S 760

is a stratified program. Proof Let us start with the following two observations: 1. Since W F S satisfies relevance [38], then S ⊆ rel rules(P, g). Hence, if S is decomposable in S1 ∪ · · · ∪ Sn such that S1 is the lower rank, g appears in the head of a clause in Sn .

765

2. Since S is minimal and S |=W F S T g, then S is acyclic with respect to negative literals. The proposition follows the previous two observations which basically check the conditions of a stratified logic program. Proposition 2 Let P and G be extended logic programs Then, the following

770

condition holds: 1. |ArgP | 6 |ArgP ∪G | Proof There are two cases to verify: 1. If LP ∩ LG = ∅, the proof is straightforward since ArgP ∪G will be ArgP ∪ ArgG

775

2. If LP ∩ LG 6= ∅, let I = LP ∩ LG and g ∈ I. Since rel rules(P ∪ G, g) = rel rules(P, g) ∪ rel rules(G, g), we can observe that if S is a minimal subset of rel rules(P, g) such that S |=W F S T g, then S is a minimal subset of rel rules(P ∪ G, g) such that S |=W F S T g. Hence, if hS, gi ∈ ArgP , then hS, gi ∈ ArgP ∪G . 32

780

Proposition 3 Let P and G be extended logic programs Then, the following condition holds: 1. |At(ArgP )| 6 |At(ArgP ∪G )| Proof The proof is straightforward, by both the fact that the number of arguments does not decrease whenever an extended logic program is extended with

785

new clauses (Proposition 2) and the definition of attacks which is based on the well-founded model of the support of each argument. Proposition 4 Let P be an extended logic program, AFP = hArgP , At(ArgP )i be the resulting argumentation framework from P and SEMGrounded be the grounded argumentation semantics [14]. AFP satisfies closure w.r.t SEMGrounded .

790

Proof Let us start by remembering that G =

S

hS,gi∈E

S and G0 =

S

a∈Output

rel rules(P, g)

(see Postulate 1). We can observe that if E ∈ SEMGrounded (AFP ) and hS, gi ∈ E, then rel rules(P, g) defines a stratified logic program. Hence, since SEMGrounded (AFP ) infers only one extension, then G = G0 . Hence the proposition follows the fact that WFS satisfies relevance.

795

Proposition 5 Let P be an extended logic program, AFP = hArgP , At(ArgP )i be the resulting argumentation framework from P. AFP satisfies local closure w.r.t SEMstable , SEMpref erred and SEMpref erred [14]. Proof The proposition follows by the fact that WFS satisfies relevance. Proposition 6 Let P be an extended logic program, AFP = hArgP , At(ArgP )i

800

be the resulting argumentation framework from P and SEMArg be an argumentation semantics. If SEMArg is a conflict-free semantics, then AFP satisfies direct consistency w.r.t. SEMArg . Proof The proposition holds to the definition of conflict-freeness by any conflict-

33

free argumentation semantics.

805

Proposition 7 Let P be an extended logic program, AFP = hArgP , At(ArgP )i be the resulting argumentation framework from P and SEMArg be an argumentation semantics. If SEMArg is a conflict-free semantics, then AFP satisfies indirect consistency w.r.t. SEMArg . Proof Let us start by remembering that G =

810

S

hS,gi∈E

S and G0 =

S

a∈Output

rel rules(P, g)

0

(see Postulate 1). Since SEMArg is conflict-free semantics, G and G will infer consistent models.

Appendix C. Implemented Algorithms for Argument Building In this section, we introduce the algorithms oriented to implement our argumentation engine approach. 815

By considering some of the definitions described in the previous sections, Algorithm 1 will generate a set of arguments w.r.t. a given atom. The input of this algorithm is an atom which is the epicenter of the inference and an extended logic program, described as follows: a combinatorial process for obtaining candidate clauses in line 5 is performed. Even when, in the worst case, the time

820

complexity for building a power set is Θ(2n ), the set of relevant rules decreases the number of combinations, rather than performing this process for the entire program. Let us note that line 6 is the principal WFS evaluation of the algorithm. In this line, those sets of clauses whose truth value allow us to infer true (T ) conclusions are selected, and only those atoms with this characteristic

825

will take part in the inference process. As a final process in our proposed main algorithm, a recurrent method for building a set of arguments using minimal supports is described in line 7. The relevant clauses search is one of the key components in our approach. Let us introduce an auxiliary algorithm which returns a set of relevant clauses

830

following Definition 5. This algorithm is presented in Algorithm 2. Let us briefly describe some important aspects of Algorithm 2. In line 4, the

34

Algorithm 1: ArgumentBuilder(a, P ) input : a ∈ LP , P output: set of arguments inferring a 1

Let S and Arguments be empty list stacks

2

Let W F S(S) = hT, F i return the well-founded model with respect to S.

3

Let rel rules(P, X) be a function returning relevant clauses for X w.r.t. P /* Definition 5 */

4

Let Minimal set(S) be a function returning the minimal set w.r.t. a set S.

5

S P := 2rel rules(a,P )

6

S W F S = {S | a ∈ T, such that S ∈ S P , W F S(S) = hT, F i and @ c ∈ LS such as{c, ¬c} ∈ T }

7 8 9 10

foreach S ∗ ∈ Minimal set(S W F S ) do

ArgumentsS = {hS ∗ , ai} ∪ ArgumentsS

end Returns ArgumentsS

Algorithm 2: RelevantRules(P, X) input : Atom X, Program P output: Set of relevant rules of P 1

Let rel rules be an empty list stack

2

Let connected components be an empty list stack

3

Let dependenciesOf(X) be a function for capturing all the dependencies of atom X

4

connected components = dependenciesOf(X)

5

foreach clause ∈ connected components do

6

rel rules = rel rules ∪ clause

7

end

8

Returns rel rules(P, X)

35

function dependenciesOf(X) returns all the connected component parts of a given atom X. This procedure is highly dependent on a search implementation. We propose a Depth-first search algorithm [57] applied to P , by considering P 835

as a graph structure. As a final remark, let us recall that a Depth-first search has a linear complexity (Θ(nodes + edges)). Appendix C.0.1. Algorithm for Counterarguments By considering the rebut and undercut attack relationships as introduced in Definition 6, we follow linear searches [57] through the Arguments stack, in

840

order to infer rebuts and undercuts. We use the relevance principle for reducing the linear search space, which is performed into the stack of arguments inferring related atoms, using the inferred atom (from the WFS evaluation) as a key. Let us describe this method as follows: Algorithm 3: RebutSearch input : Algorithms stack, atom output: Rebut attack 1

Let Head(Arg) be a function returning the head of an argument

2

Let ConnectedComp (S) be a function returning the connected components of a set S

3

Let Rebuts be an empty stack

4

x = ConnectedComp (Arguments)

/* Obtains a set of sets of

connected clauses */ 5

while x 6= Null AN D Head(x) 6= atom do

6

x = x.next

7

if x = ¬atom then Rebuts = Rebuts ∪ x

8 9

end

10

end

11

return Rebuts

36

Let us highlight a few details about Algorithm 3. The time complexity of 845

a linear search depends on the entire stack, which, in our approach, is Θ(n) in the worst case. On this point, it is important to highlight the importance of Line 4 reducing the linear search space. In the case of an undercut relationship, we also use the relevant clauses, but we made a linear search in the support set of each argument.

37

850

Appendix D. Full Example Generating Arguments Full Argument Building Process Atom

c

Relevant Rules

rel rules(P, c) = { {∅} , {c ← not d}} |{z} | {z } S0

WFS Evaluation Arguments

S1

W F S T (S0 ) = {}

W F S T (S1 ) = {c} Arg1 = h{c ← not d}, ci | {z } S1

Atom

d

Relevant Rules

rel rules(P, d) = { {∅} , {d ← not c}} |{z} | {z } S0

WFS Evaluation Arguments

S1

W F S T (S0 ) = {}

W F S T (S1 ) = {d} Arg2 = h{d ← not c}, di {z } | S1

Atom

a

Relevant Rules

rel rules(P, a) = { {∅} , {a ← n}, {a ← >}, |{z} | {z } | {z } S0

S1

S2

{n ← >}, {a ← n; a ← >}, {a ← n; n ← >}, {z } | {z } | {z } | S3

S4

S5

{a ← >; n ← >}, {a ← n; a ← >; n ← >}} | {z } | {z } S6

S7

W F S T (S0 ) = {} W F S T (S1 ) = {}

W F S T (S2 ) = {a} WFS Evaluation

W F S T (S3 ) = {n} W F S T (S4 ) = {a}

W F S T (S5 ) = {a, n} W F S T (S6 ) = {a, n} W F S T (S7 ) = {a, n} Arguments

Arg3 = h{a ← >}, ai | {z } S2

Arg4 = h{a ← n; n ← >}, ai | {z } S5

38

Full Argument Building Process (continuation) Atom Relevant Rules

n rel rules(P, n) = { {∅} , {n ← >}, {a ← n; n ← >}, |{z} | {z } | {z } S0

S1

S2

{a ← >; n ← >}, {a ← n; a ← >; n ← >}} {z } | {z } | S3

S4

W F S T (S0 ) = {}

W F S T (S1 ) = {n} WFS Evaluation

W F S T (S2 ) = {n, a} W F S T (S3 ) = {n, a} W F S T (S4 ) = {n, a}

Arguments

Arg5 = h{n ← >}, ni | {z } S1

39

Appendix E. Use Case We introduce here a use case for our argument engine, describing the argu855

ment building process. This use case is part of the Anna scenario context, formally introduced in [53]. Monitoring physical activity patterns of Anna. Anna agrees with therapists to use an activity monitoring argument-based system to record her physical activity patterns. This system is installed on her mobile phone. In this setting, the system obtains a register of her location and

860

locomotion activities over a period of time. Each observation is captured by an extended logic program clause. Since the information obtained from the sensors is pervaded by vagueness, an observation with uncertainty data will be described by using negation as failure, like: She0 s at Rest ← not She0 s Jogging, which intuitively illustrates that the system does not have confidence that Anna is

865

jogging, hence she is resting.

Figure E.6: Anna use case

Let us define the set of arguments, first establishing the knowledge epicenter in the atom She0 s Jogging. By considering the relevant rules of She0 s Jogging, we find that: The relevant rules for this atom are:

870

The WFS evaluation (W F S(S) = hT, F i) for sets S0 , S1 , S2 , S3 is: W F S(S0 ) = {∅, ∅} 40

W F S(S1 ) = {She0 s Jogging, She0 s at Rest}

W F S(S2 ) = {She0 s at Rest, She0 s Jogging} W F S(S3 ) = {∅, ∅} 875

From this evaluation we can obtain the following set of arguments: Arg1 = h{S1 }, She0 s Joggingi Arg2 = h{S2 }, She0 s at Resti Following the same procedure, the entire set of built arguments for the pro880

gram P is: Arg1 = h{S1 }, She0 s Joggingi Arg2 = h{S2 }, She0 s at Resti

Arg3 = h{S4 }, She0 s walkingi

Arg4 = h{S5 }, She0 s leavingHomei Arg5 = h{S6 }, She0 s walkingi

885

Where: S4 = {She0 s walking ← She0 s leavingHome; She0 s leavingHome ← >}

S5 = {She0 s leavingHome ← >} and S6 = {She0 s walking ← >}

890

As Arg5 has a minimal support for She0 s walking than Arg3 , consequently Arg3 is discarded. The argument inference process is finished when one of the built arguments is selected; in the Anna use case, an agent sends a notification to her mobile phone, depending on the observation registered by the sensors 895

(see [53]). In these kinds of scenarios, we believe that our argumentation tool could have an important practical utility. ————————————————————————————-

41

74

II

Reasoning about Human Activities: an Argumentative Approach Juan Carlos NIEVES a,1 , Esteban GUERRERO a and Helena LINDGREN a a Department of Computing Science, Ume˚ a University, Sweden Abstract. Recognizing and supporting human activities is an important challenge for ambient assisted living. In this paper we introduce a novel argumentation-based approach for dealing with human activity recognition. By considering a model of the world and a set of observations of the world, hypothetical fragments of activities are built. The hypothetical fragments of activities will be goal-oriented actions and they will be considered defeasible. Therefore we consider extension-based argumentation semantics for local selection of hypothetical fragments of activities. By considering degrees of fulfillment of activities and local selection, a global selection of hypothetical fragments of the activities is defined. Therefore, we can make explicit statements about why one hypothetical activity was performed. Keywords. Human Activities Recognition, Argumentation Reasoning

1. Introduction Human activity recognition is a challenging field which goes in a different direction compared to planning methods [10]. For instance, multiple hypotheses are possible regarding the intentions of an observer agent. Knowing a user’s activities and goals can significantly improve the effectiveness of the services of an intelligent system. For instance, if a smart home (or a human-aware robot) is providing support to elderly people, it can provide adequate assistive services at the opportune moment, e.g., if a person has mental impairments, the smart home (or the robot) can provide reminders, dynamic tutorials of basic activities such as cooking, etc. Human activity can be understandable in terms of Activity Theory (AT). AT is a theoretical framework for studying different forms of human praxis [3]. AT provides a dynamic model of human agent activity that integrates both motivational factors (objectives) sprung from needs, goal-directed actions (cognitive processes) as well as unconscious acts or behavior (operations). In AT, an activity is defined by a motive or an objective and is executed by means of a set of goal-directed actions. In a recent work, a systemic-structural analysis of a human activity has been explored [2]. In this analysis, a human activity is carried out through actions, realizing the objective of an activity and generating an outcome. These actions are governed by con1 Correspondence to: Department of Computing Science, Ume˚ a University, SE-901 87, Ume˚a, Sweden; E-mail: J.C. Nieves: [email protected]; Esteban Guerrero: [email protected]; Helena Lindgren: [email protected]

scious goals of the subject. A goal reflects the result of an action; hence, the sum of goals reflects the overall objective of an activity. In this setting, an activity can regarded as a set of goals. Currently there are few reasoning methods which could be considered suitable for reasoning about human activities. On the one hand, there are some works in terms on action specification languages [1,7,4]. In these works, the atomic relation between conscious goals and actions which is suggested by activity theory is not explicit; therefore, the explanation of both fulfillment and non-fulfilment of activities in terms of goals and actions is not straightforward. On the other hand, in the context of argumentation reasoning, there are less proposals for reasoning about activities [10]. However, we can highlight the work of Konolige and Pollack [9] for plan recognition. In this approach, the authors argue for a bottom-up approach based on argumentation and two levels of selection of plan-fragments. Against this background, we introduce a novel bottom-up approach for activity recognition based on argumentation reasoning. More accurately, in order to define the concept of hypothetic fragments of activities, we follow the ideas of activity theory which suggests that actions are motivated by needs. The hypothetical fragments of activities will be basically hypothesis about small pieces of activities. In order to select hypothetic fragments of activities which suggest evidence of the fulfillment or non-fulfilment of particular activities, we follow a strategy similar to the one explored by Konolige and Pollack which is based on two selections: Local selection and Global Selection [9]. 1.- Local Selection: The local selection is the first step in the selection of hypothetic fragments of activities. In the local selection, the hypothetic fragments of activities are considered defeasible; therefore, we use argumentation semantics [5,6] for selecting a set of sets of hypothetic fragments of activities. This means that basically the local selection leads with the defeasible information which is present in the hypothetic fragments of activities. 2.- Global Selection: By considering the sets of hypothetical fragments of activities suggested by the local section, the global selection defines rules for determining which sets of hypothetical fragments of activities could suggest evidence about the fulfillment or non-fulfilment of some particular activities. To this end, the global selection defines different degrees of fulfilment and non-fulfilment of activities. As it was pointed out by Konolige and Pollack, both the local selection and global selection are not easy processes since we will must look for coherent sets of fragments of activities. By coherent, we mean that the union of fragments of activities could suggest the fulfillment of particular activities. Given that Dung’s argumentation semantics are particular relevant in argumentation reasoning, we also explore how the global selection of hypothetical fragments of activities is affected by considering Dung’s argumentation semantics at local level selection. In general terms part of the contributions of this paper are: • The integration of activity theory and argumentation theory. To the best of our knowledge, this is the first work which tries to achieve this goal. • An extension of Dung’s argumentation approach for capturing activity theory. • To define a new approach for human activity recognition for smart environments.

• The suggested approach recognizes a human activity and it is also able to argue why a given human activity is achieved or not. • The behavior of generic argumentation semantics and Dung’s argumentation semantics w.r.t. activity recognition is studied The rest of the paper is structured as follows: In Section 2, a basic introduction to Dung’s argumentation semantics is presented. In Section 3, the concepts of activity framework and hypothetical fragment of an activity are introduced. In Section 4, the local selection of hypothetical fragments of activities is described. To this end, an attack relation between hypothetical fragments of activities is defined. In Section 5, the global selection of hypothetical fragments of activities is defined. To this end, some rules of fulfilment of activities are introduced. In Section 6, the behavior of the Dung’s argumentation semantics w.r.t. activity recognition is studied. In the last section, an outline of our conclusions are presented. Due to lack of space, we omit the formal proofs of the theoretical results. 2. Background In this section, we introduce some basic concepts of argumentation semantics. We start defining some basic concepts of Dung’s argumentation approach. The first one is an argumentation framework. An argumentation framework captures the relationships between arguments. Definition 1 [5] An argumentation framework is a pair AF := ⟨AR, attacks⟩, where AR is a finite set of arguments, and attacks is a binary relation on AR, i.e.attacks ⊆ AR × AR. We say that a attacks b (or b is attacked by a) if attacks(a, b) holds. Similarly, we say that a set S of arguments attacks b (or b is attacked by S) if b is attacked by an argument in S. Let us observe that an argumentation framework is a simple structure which captures the conflicts of a given set of arguments. In order to select coherent points of views from a set of conflicts of arguments, Dung introduced a set of patterns of selection of arguments. These patterns of selection of arguments were called argumentation semantics. Dung defined his argumentation semantics based on the basic concept of admissible set: Definition 2 [5] • A set S of arguments is said to be conflict-free if there are no arguments a, b in S such that a attacks b. • An argument a ∈ AR is acceptable with respect to a set S of arguments if and only if for each argument b ∈ AR: If b attacks a then b is attacked by S. • A conflict-free set of arguments S is admissible if and only if each argument in S is acceptable w.r.t. S. By considering the concept of admissible set, Dung et al. introduced five basic argumentation semantics: the grounded, stable, preferred, complete and ideal semantics [5,6]. Even though all of them are based on admissible sets, each of them represents a different pattern of selection of arguments.

Definition 3 [5,6] Let AF := ⟨AR, attacks⟩ be an argumentation framework. An admissible set of argument S ⊆ AR is:

• a stable extension if and only if S attacks each argument which does not belong to S. • a preferred extension if and only if S is a maximal (w.r.t. inclusion) admissible set of AF . • a complete extension if and only if each argument, which is acceptable with respect to S, belongs to S. • a grounded extension if and only if it is a minimal (w.r.t. inclusion) complete extension. • an ideal extension if and only if it is contained in every preferred set of AF.

SEMx (AF) denotes the set of extensions of the argumentation framework AF with respect the argumentation semantics x such that x ∈ {s, p, c, g, i} where s stands for stable, p for preferred, c for complete, g for grounded and i for ideal. SEM denotes a generic AF argumentation semantics such that SEM is a function from AF to 22 . Given that the set of ideal sets of an argumentation framework defines a total ordered set (w.r.t. inclusion), the maximal ideal set is usually the interesting ideal set to infer from an argumentation framework. Hence, SEMi (AF) denotes the maximal ideal set of AF. 3. Activity Argumentation Frameworks In this section, we will introduce the concept of an activity framework. An activity framework will define all the components for building hypothetical fragments of activities. These fragments of activities will define hypotheses about activities. We assume that the reader is familiar with basic concepts of classical logic. The reader can find a good introduction to classical logic in [11]. In what follows, ⊢ denotes classical inference and ≡ denotes logical equivalence. An activity framework will follow a structure of cognitive state of an agent namely beliefs, desires and intentions: Definition 4 (An Activity Framework) An activity framework ActF is a tuple of the form ⟨T, HA , G , O, Acts⟩ in which:

• T is a propositional theory. LT denotes the set of atoms which appears in T . • HA = {d1 , . . . , dn } is a set of atoms such that HA ⊆ LT . HA denotes the set of hypothetical actions which an agent can perform in a world. • G = {g1 , . . . , gn } is a set of atoms such that G ⊆ LT . G denotes a set of goals of an agent. • O = {o1 , . . . , on } is a set of atoms such that O ⊆ LT . O denotes a set of observation from a world. • Acts ⊂ 2G . Acts denotes a set of activities. We are assuming that a set of goals defines an activity.

Given an activity framework, one can build small pieces of knowledge which give hypothetical evidence of the achievement of a given goal by considering a set of be-

lieves (a set of proportional formulas), a hypothetical action and a set of observations of the world. These small pieces of knowledge will be called hypothetical fragments of activities: Definition 5 (A Hypothetical Fragment of an Activity) Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework. A hypothetical fragment of an activity is of the form ⟨S, O′ , a, g⟩ such that: • • • •

S ⊆ T , O′ ⊆ O, a ∈ HA and g ∈ G S ∪ O′ ∪ {a} is consistent. S ∪ O′ ∪ {a} ⊢ g S and O′ are minimal w.r.t. set inclusion.

Let us denote by HFActF the set of hypothetical fragments which we can construct from ActF. Let us observe that a hypothetical fragment of an activity is basically a goal-oriented action which takes as input observations of the world. From an intuitive point of view, the construction of hypothetical fragments of activities represents the process of building hypotheses about the fulfillment of some possible activities. Since the hypothetical fragments of the activities are based on hypothetical actions, the hypothetical fragments of activities are defeasible. In order to deal with the defeasible information which is present in the hypothetical fragments of activities, we will follow a defeasible reasoning process based on attack relations between the hypothetical fragments of the activities and argumentation semantics. These two elements will be the core for the local selection ( the first selection) of the hypothetical fragments of the activities

4. Local Selection of Hypothetical Fragments of Activities The selection of sets of hypothetical fragments of activities is managed by two steps: Local Selection and Global Selection. The aim of the local selection is to deal with the defeasible information which is presented in the hypothetical fragments of activities. Hence, we will follow an argumentation reasoning approach for selecting sets of hypothetical fragments of activities which could suggest potential fulfillment of activities. In particular, we will use argumentation semantics for selecting sets of hypothetical fragments of activities. To this end, we start defining an attack relation between hypothetical fragments of activities. Definition 6 (Attack relation) Let ActF be an activity framework and F1 , F2 ∈ HFActF such that F1 = ⟨S1 , O′1 , a1 , g1 ⟩, F2 = ⟨S2 , O′2 , a2 , g2 ⟩. F1 attacks F2 if one of the following conditions hold: • ∃x ∈ S2 such that x ≡ ¬g1 . • g2 ≡ ¬g1 AttHFActF denotes the set of attacks which occurs between hypothetical fragments of activities which belong to HFActF .

By regarding hypothetical fragments of activities as arguments, Definition 6 basically is defining an attack relation between arguments. Therefore, we can use argumentation semantics for selecting sets of hypothetical fragments of activities. To this end, let us define the concept of argumentation activity framework as follows: Definition 7 Let ActF be an activity framework. An activity argumentation framework AAF with respect to ActF is of the form: AAF = ⟨ActF, HFActF , AttHFActF ⟩. By abusing of notation, an argumentation semantics SEM, as was defined in Section 2, will infers a set of sets of hypothetical fragments of activities from an activity argumentation framework. In this setting, activity argumentation frameworks are regarded as argumentation frameworks and hypothetical fragments of activities as atomic arguments. Therefore, an argumentation semantics SEM will define the local selection (initial selection) of hypothetical fragments of activities. 5. Global Selection of Hypothetical Fragments of an Activity Selecting hypothetical fragments of activities by considering argumentation semantics is only one of the steps of activity recognition. An argumentation semantics can only suggest multiple competing sets of hypothetical fragments of activities which could suggest the fulfillment of some activities. Therefore, we require a global selection of hypothetical fragments of activities. By global selection, we mean a selection able to: • suggest degrees of both fulfillment and non-fulfillment of activities; moreover, • suggest evidence for believing about the fulfillment of activities.

Given that a hypothetical fragment of an activity always has a goal, a set of hypothetical fragments of activities can be regarded as a set of goals. To this end, let us define the following notation: Given a set of hypothetical fragments of activities E, E G is defined as follows: E G = {g|⟨S, O′ , a, g⟩ ∈ E}. By considering that a set of hypothetical fragments of activities can be regarded as a set of goals, the status of an activity is defined as follows: Definition 8 (Status of Activities) Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF and SEM be an argumentation semantics. An activity Act ∈ Acts is:

• achieved iff Act ⊆ E G for all E ∈ SEM(AAF). • partially-achieved iff ∃E ∈ SEM(AAF) such that Act ⊆ E G and ∃E ′ ∈ SEM(AAF) such that act * E ′G • null-achieved iff for all E ∈ SEM(AAF), Act * E G

It is important to observe that an extension E ∈ SEM(AAF) is suggesting fragments of activities which argue why a particular activity is fulfilled. One can observe that there are preservations of status between activities. For instance, one can show that any sub-activity from an achieved activity is achieved, the intersection of two achieved activities is achieved and the union of two achieved activities is achieved.

Proposition 1 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF and SEM be an argumentation semantics. a) If Act1 , Act2 ∈ Acts such Act2 ⊆ Act1 and Act1 is achieved w.r.t. SEM(AAF) then Act2 is achieved w.r.t. SEM(AAF). b) If Act1 , Act2 , Act3 ∈ Acts such Act1 ∩ Act2 = Act3 , and Act1 , Act2 are achieved w.r.t. SEM(AAF) then Act3 is achieved w.r.t. SEM(AAF). b) If Act1 , Act2 , Act3 ∈ Acts such Act1 ∪ Act2 = Act3 , and Act1 , Act2 are achieved w.r.t. SEM(AAF) then Act3 is achieved w.r.t. SEM(AAF). Similar to the case of achieved activities, one can see that any activity which contains null-achieved sub-activities are null-achieved. Proposition 2 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF and SEM be an argumentation semantics. a) If Act1 , Act2 ∈ Acts such Act2 ⊆ Act1 and Act2 is null-achieved w.r.t. SEM(AAF) then Act1 is null-achieved w.r.t. SEM(AAF). b) If Act1 , Act2 , Act3 ∈ Acts such Act1 ∪ Act2 = Act3 and Act1 , Act2 are null-achieved w.r.t. SEM(AAF) then Act3 is null-achieved w.r.t. SEM(AAF). By considering the number of goals of each activity, one can define different degrees of achievement w.r.t. each activity. Indeed, one can define a degree of achievement and a degree of non-achievement. Definition 9 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF, SEM be an argumentation semantics and Act1 ∈ Acts such that Act2 ⊆ Act1 .

• Act1 is (i/n)-achieved if Act2 is achieved w.r.t. SEM(AAF), i = |Act2 | and n = |Act1 |. • Act1 is (1−i/n)-null-achieved if Act2 is achieved w.r.t. SEM(AAF), i = |Act2 | and n = |Act1 |. • Act1 is (i/n)-hard-null-achieved if for all E ∈ SEM(AAF), Act2 ∩ E G = 0, / i= |Act2 | and n = |Act1 |.

From the degrees of achieving introduced by Definition 9, one can identify relationships between them. Proposition 3 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF and SEM be an argumentation semantics. a) If Act ∈ Acts is achieved, then Act is (1)-achieved otherwise Act is (i)-acceptable such that i < 1. b) If Act ∈ Acts is partially-achieved, then Act is (i)-achieved such that i < 1. c) If Act ∈ Acts is null-achieved, then Act is (1)-null-achieved.

Proposition 4 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF and SEM be an argumentation semantics. a) If Act ∈ Acts is 0-achieved, then Act is null-achieved. b) If Act ∈ Acts is 0-null-achieved, then Act is achieved. By considering the number of extensions that make a given activity partiallyachieved, one can define a preference relations between partially-achieved activities. Definition 10 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF, SEM be an argumentation semantics, Act1 , Act2 ∈ Acts such that Act1 , Act2 are partiallyachieved activities. The preference relation ≽n between partially-achieved activities is defined as follows: Act2 ≽n Act1 if and only if |E (Act2 , SEM(AFF))| ≥ |E (Act1 , SEM(AFF))| where E (Act, SEM(AFF)) = {E|E ∈ SEM(AFF) and Act ⊆ E}. 6. The status of an activity by considering Dung’s semantics In the previous section, we have explored the status of activities w.r.t. no-particular argumentation semantics. In this section, we will identify some basic relations w.r.t. the status of an activity by considering different argumentation semantics based on admissible sets. As we saw in Section 2, Dung et al. introduced five argumentation semantics. All of them are based on the concept of an admissible set. These semantics can be split into two groups: skeptical semantics and credulous semantics. On one hand, the grounded and ideal semantics are considered skeptical semantics. On the other hand, the stable, preferred and complete semantics are considered credulous semantics. Given that the skeptical semantics only identify a unique extension from a given argumentation framework, the grounded and ideal semantics do not identify partiallyachieved activities. Proposition 5 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF, SEM be an argumentation semantics and NSEM (AAF) = {A|A ∈ Acts and A is partiallyachieved w.r.t. SEM(AAF)}. a) NSEMg (AAF) = 0/ b) NSEMi (AAF) = 0/ It is known that the maximal ideal set is a superset of the grounded semantics [6]. Hence, the status of an achieved activity or a null-achieved activity w.r.t. the grounded semantics is preserved w.r.t. the ideal semantics. Proposition 6 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF and A ∈ Acts.

a) If A is achieved w.r.t. SEMg (AAF) then A is achieved w.r.t. SEMi (AFF). b) If A is null-achieved w.r.t. SEMg (AAF) then A is null-achieved w.r.t. SEMi (AFF). By the definition of the ideal semantics, it is known that an ideal extension is contained in every preferred extension. Hence, one can identify the following property w.r.t. achieved activities and ideal semantics. Proposition 7 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF and A ∈ Acts. If A is achieved w.r.t. SEMi (AAF) then A is achieved w.r.t. SEM p (AAF). Dung showed that every stable extension is a preferred extension but not vice versa [5]. This property suggests the following properties between activities and these argumentation semantics: Proposition 8 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF and A ∈ Acts. a) If A is achieved w.r.t. SEM p (AAF) and SEMs (AAF) ̸= 0/ then A is achieved w.r.t. SEMs (AAF). b) If A is partially-achieved w.r.t. SEMs (AAF) then A is partially-achieved w.r.t. SEM p (AAF). c) If A is null-achieved w.r.t. SEM p (AAF) and SEMs (AAF) ̸= 0/ then A is null-achieved w.r.t. SEMs (AAF). Given that the grounded semantics is exactly the intersection of the complete extensions and the grounded extension is a subset of the intersection of the preferred extensions, one can identify the following properties. Proposition 9 Let ActF = ⟨T, HA , G , O, Acts⟩ be an activity framework, AAF = ⟨ActF, HFActF , AttHFActF ⟩ be an activity argumentation framework with respect to ActF and A ∈ Acts. a) If A is achieved w.r.t. SEMg (AAF) iff A is achieved w.r.t. SEMc (AFF). b) If A is achieved w.r.t. SEMg (AAF) then A is achieved w.r.t. SEM p (AFF). c) If A is achieved w.r.t. SEMc (AAF) then A is achieved w.r.t. SEM p (AFF). 7. Conclusions In this paper we have presented a novel bottom-up approach to human activity recognition. This approach takes as starting point activity theory which argues for goal-oriented actions which are motivated by needs. In our approach the general problem of activity recognition is captured by the so called activity frameworks which have as input a predefined set of activities in terms of sets of goals. In order to recognize activities, we build hypothetic fragments of activities from a given activity framework. We defined a calculus of attacks between hypothetic fragments of activities in order to deal with the defeasible information which is present in the hy-

pothetic fragments of activities. Therefore, the selection of hypothetic fragments of activities is based on two selections: a local selection and a global selection. Both the local selection and global selection are not easy processes since we will must look for coherent sets of fragments of activities. We have shown that by considering argumentation semantics for the local selection, we can define different degrees of fulfilment and non-fulfilment of activities at global selection level. Let us observe that these degrees of fulfilment and non-fulfilment define a kind of error-measurement. It is worth mentioning that authors such as Kautz [8] has pointed out the relevance of defining error-measurements in plan recognition. Hence, it is quite obvious that considering error-measurements is also relevant in activity recognition. We have identified different relations w.r.t. the status of an activity by considering different argumentation semantics. These relations define different strategies for implementing our approach. As part of our future work, we consider an implementation of our approach in order to validate the suggested approach with real scenarios. Due to lack of space, we have not presented examples. Acknowledgment This research has been partially supported by VINNOVA (The Swedish Governmental Agency for Innovation Systems) and the Swedish Brainpower. References [1] C. Baral and M. Gelfond. Reasoning about Intended Actions. In AAAI, pages 689–694. AAAI Press / The MIT Press, 2005. [2] G. Bedny and W. Karwowski. Introduction to applied and systemic-structural activity theory. In HumanComputer Interaction and Operators Performance: Optimizing Work Design with Activity Theory, page 464. CRC Press Taylor & Francis Group, Boca Raton, Florida, 2011. [3] O. W. Bertelsen and S. Bø dker. Activity Theory. In J. M. Carrol, editor, HCI Models, Theories, and Frameworks: Toward an Interdisciplinary Science, chapter 11, pages 291–324. Morgan Kaufmann, 2003. [4] J. Blount and M. Gelfond. Reasoning about the intentions of agents. In Logic Programs, Norms and Action, volume 7360 of Lecture Notes in Computer Science, pages 147–171. Springer, 2012. [5] P. M. Dung. On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games. Artificial Intelligence, 77(2):321–358, 1995. [6] P. M. Dung, P. Mancarella, and F. Toni. Computing ideal sceptical argumentation. Artificial Intelligence, 171(10-15):642–674, 2007. [7] A. Gabaldon. Activity recognition with intended actions. In IJCAI, pages 1696–1701, 2009. [8] H. Kautz. Reasoning About Plans, chapter .A Formal Theory of Plan Recognition and its Implementation, pages 69–126. Morgan Kaufmann Publishers, 1991. [9] K. Konolige and M. E. Pollack. Ascribing plans to agents. In Proceedings of the 11th International Joint Conference on Artificial Intelligence (IJCAI). Detroit, MI, USA, pages 924–930. Morgan Kaufmann, 1989. [10] F. Sadri. Ambient Intelligence and Smart Environments: Threds and Perspectives, chapter Logic-Based Approaches to Intention Recognition, pages 346–375. IGI Gobal, 2011. [11] D. van Dalen. Logic and structure. Springer-Verlag, Berlin, 3rd., aumented edition edition, 1994.

III

ALI, an Ambient Assisted Living System for Supporting Behavior Change Esteban Guerrero, Helena Lindgren, and Juan Carlos Nieves (esteban, helena, jcnieves)@cs.umu.se Ume˚ a University SE-901 87, Ume˚ a, Sweden

Abstract. We introduce ALI, a modular agent-based ambient assisted living system, for supporting behavior change in mildly depressed and socially isolated individuals. The aim for this interdisciplinary work is empowering an individual through positive feedback and through encouraging the individual to follow recommendations formulated by the individual him or herself, in order to promote behavior change. ALI is an initiative for recognizing and explaining human behavior, providing argument-based explanations for goal-based activities, using a knowledge repository (ACKTUS), and agents for capturing situations. We introduce the term situation for defining an atomic observation of the world, obtained by ALI instantiated in a exchangeable format. ALI is exemplified in the paper from the perspective of a mildly depressed person with social withdrawal and mild cognitive impairment, and is implemented as a mobile application application.

1

Introduction

Becoming isolated in one’s home may cause a sense of loneliness and depressive mood. Depression in turn, is typically observed by a social withdrawal and indifference for social contacts and activities, which used to be important to the individual. Moreover, a noticeable decrease in psychomotor activity is also seen. This circle of decrease in activity and increase of indifference and depressed mood is very difficult for the individual to change, mainly because the disease takes away exactly this: the armor necessary to change the situation. On the other hand, when the withdrawal pattern of behavior is challenged, and the individual becomes exposed to a positive environment disrupting the destructive depressed mood, this may push the spiral in a constructive direction towards a more empowered behavior. Activity recognition is determinant to the success of context-aware personalized systems in a number of pervasive computing and smart environments. Different approaches have been developed based on embedded mobile sensors in order to sensing and recognize: community or social interactions [1], physical motion [2], eating patterns [3], mild cognitive impairments (MCI) [4–6]. The contribution of this paper is the argumentation-based multi-agent system: ALI (Assisted LIving system), developed for providing support for behavior

change in individuals with mild depression and social withdrawal. An agentbased activity recognition system, which bases its interpretations of the individual’s activities on the sensors of a mobile phone, can initiate argument-dialogues with other agents, in order to obtain strong evidence of behavior change. The purpose of the dialogues is to reason about suitable motivational arguments, obtaining explanations for his/her behavior and at the same time, providing to the individual with advices in order to achieve a behavior change. From a therapeutic perspective, it is important to detect changes in activity levels in fragile individuals. If there is a decrease in activity level, this may indicate that a depression is present, which may require treatment. On the other hand, increase in activity may indicate a positive response to treatment, or that treatment needs adjustments. This paper presents ALI exemplified in a case scenario of an individual living almost all her time in an Ambient Assisted Living (AAL) environment. ALI fulfills the following five design requirements: 1.- it is a non-intrusive pattern disorder recognition alternative; 2.- it is scalable and flexible; 3.- deals with uncertain and incomplete information from sensors with no data training; 4.the daily living activity recommendation is monitored by a therapist; and 5.- its knowledge foundation in the reasoning and decision making process is supported by a set of knowledge repositories ACKTUS [7]. The rest of the paper is divided as follows. In the next section a use case scenario is presented, which will be a running example for the rest of the paper. Section 3 presents the ALI architecture explaining all its components. Finally, Section 4 provides some conclusions and implications for future work.

2

Use Case and Application Context

In order to exemplify each architectural component and data exchange we introduce in this section a use case and the context in which ALI runs. A woman Rut who is 84 years old, is diagnosed with MCI [8]. She is also mildly depressed and feels isolated due to difficulties walking outside with her walking aid. She is very worried that she might fall. She also has some difficulties when performing daily activities due to her pain condition. In addition, assessments indicate that she has some sleeping problems. In this scenario, a therapist evaluates together with Rut her main daily activities in order to improve her activity performance towards a more satisfactory level. Rut prioritizes her activities, assigning importance to the activities she wants to be able to accomplish (Figure 1). In our scenario the health care personnel offer Rut a method to monitor her sleeping patterns and to investigate how her functional difficulties affect her social life. While Rut may be provided a better view of her activities over time, the therapists are given means to synthesize evidence collected by using interviews and questionnaires, with evidence obtained by an activity recognition system. In addition, based on the activity pattern and goal and priority settings the system

may propose and motivate different actions to Rut in suitable situations, which may help Rut in her decision making processes about activities in daily life. Rut agrees to test the method, which means to carry along her mobile phone whenever she is inside or outside of her home. She is also instructed to put the phone under her pillow when she goes to sleep in her bed. This for the purpose to detect her movements and sounds during night-time. In this scenario, ALI obtains and analyzes data from the Rut context using her mobile phone, which has different embedded sensors (accelerometer, gyroscope and audio). At the same time, ALI exchanges information with ACKTUS repositories in order to provide extra evidence to her therapists about potential behavior changes and her symptoms evolution.

Fig. 1. The goal setting activity illustrated through screenshots of Rut’s I-Help application, her evidence in the ACKTUS-care view and in the ACKTUS-dementia view.

ACKTUS is a semantic web-based platform for developing knowledge-based support applications. ACKTUS utilizes a core ontology which has as part an extended version of the Argument Interchange Format (AIF) developed for exchanging arguments on the World Wide Web [9]. Key nodes in the ontology are the information nodes called i-nodes, carrying defeasible facts, and the schemenodes (s-nodes), which combine i-nodes into structures which carry procedural knowledge. Domain professionals use the ACKTUS editors to model domain knowledge and to some extent also the interaction with the knowledge. Systems may then extract arguments into different formats for reasoning. Key properties of the i-nodes are that 1) they are identified through a concept, which is typically retrieved from medical terminologies and 2) they carry at least one

value, which gives the uncertainty level of the piece of information. A node may also carry additional values characterizing the observed phenomenon with severity, distribution, etc. Key property of a s-node is the knowledge source it implements, modeled as an argumentation scheme. To each knowledge domain a domain repository is built, and a user case repository. In the case when a health professional assesses a patient, information is also stored in a patient repository. In our scenario Rut is using I-Help as a support application. I-Help is an application for performing an initial assessment of an individual, providing critical question to be answered by the individual. I-Help also has some other functionalities obtained from ACKTUS-care [7]. The therapist uses ACKTUS-care and the physician uses DMSS-W, a tool for diagnosing cognitive disorders. An example of how the goal setting and dialogues generating recommendations are accomplished is given in Figure 1.

3

ALI: a Mobile Activity Recognition System in an AAL Environment

In this section we introduce ALI as a modular agent-oriented architecture (Figure 2). We define all the interactions among different agent-based components and data interactions using the Rut scenario analyzed in Section 2. The component interactions are defined in a step-by-step manner, describing all the interactions that Rut and ALI have, in order to analyze potential behavior change.

HTTP

Transactions Agent Situation Composer Agent Sampling Sensor Agent

Reasoner Agent

Recommender Agent

Transaction Agent

Tracker Agent

Authentication Registration

ALI-MOB ALI-MID

ALI-CTRL

Coach Agent Repositories

Knowledge Repositories

Domain Agent

ACKTUS

Fig. 2. ALI,the Ambient Assisted Living System and its main components

3.1

Rut Defining Her Goals

In the scenario described in Section 2, Rut is conscious about her sleeping disorders, and she decides to use ALI, which is an application running in her mobile

(Figure 3 - left). ALI offers services for analyzing her sleeping patterns as well as her social interaction patterns. When Rut presses the button “Sleep analysis” ALI performs the following tasks: – Send an activation signal to the Sampling Sensor Agent (Figure 2 in ALIMOB). The Sampling Sensor Agent starts to collect information using the appropriate sensor: • Sleeping analysis: accelerometer, gyroscope and audio sensors • Social interaction analysis: accelerometer, gyroscope, gps, network location service – Send the “sleep analysis” message with its priority to all the agents in ALICTRL, in order to perform different procedures (further down described).

Fig. 3. ALI System. Left: Activity choosing, Right: Activity recommendation message

When the user selects in the user interface an activity to analyze, this implies that it is a preferred activity. In this prototype architecture we analyze only two activities “Sleep analysis” and “Social interaction analysis”, but it is possible to use the same approach for monitoring different behavior patterns depending the context and the specific user. Once Rut has defined her goal(s), ALI-MOB and ALI-CTRL performs different tasks. Henceforth we describe all the interactions among ALI-MOB, ALICTRL and ACKTUS repositories as a transparent task, with no further details to ALI-MID, which performs authentication and log activities for all the data exchanges. 3.2

ALI and ACKTUS Interaction

There are two main data interactions between ALI modules and a knowledge repository, in our case ACKTUS.

1. Fetching person-specific information (i-nodes) from ACKTUS related to a defined goal. Each goal carries a value, defining its priority level. This interaction starts when Rut is choosing an activity to analyze. Other person-specific information from the ACKTUS user repository and domain-related knowledge obtained from the ACKTUS domain repositories, contain knowledge useful for explaining behavior and for generating recommendations about activities, also in the form of i-nodes. 2. Storing new evidence in the user case repository obtained and collected by monitoring Rut activities (Tracker Agent). An example is information about sleep disorders, performed by the Tracker Agent. This process is performed every week following a routine tailored by the therapist to Rut’s situation and needs. 3.3

Sensing and Situation Composition

Once Rut has chosen an activity and the Transaction Agent has retrieved related knowledge from ACKTUS, the Sampling Sensor Agent performs different tasks like managing sensor life cycle and getting relevant data from each specific sensor (accelerometer, gyroscope, microphone, etc.). The Sampling Sensor Agent uses a Relevant data selection sub-module (Figure 5), which obtains highest and lowest levels of different raw sensor metrics. In the case of sleep scenario illustrated in Section 2, the Relevant data selection sub-module obtains movement measures when the individual is in bed, this measure depends of the sensor sensitivity. We made different test with two different mobile phones with ALI application installed, the results showed that a naive alternative to obtain relevant data is using a high-pass filter with a cutoff value of half of the gravity value, before using the Java Android SDK. For further implementation details about the simple filter of the accelerometer sensor, we suggest follow the Java Android SDK API Guide Web page [10], where is the code implementation. A most efficient, automatic and autonomous method to filter most relevant data will be part of our future work. Each activity of an individual is monitored by the Sampling Sensor Agent, collecting data from different sensors in a defined period, which we call a situation. A situation requires be defined in terms of data stream models due to the nature of the sampling sensor process, but also it requires at the same time a formal knowledge representation format. This format must maintain a straightforward compatibility with ontologies and most known knowledge representation formats (OWL, RDF/XML, etc.). These requirements assures scalability, interoperability and adaptivity of our system. We introduce a situation representation format (Figure 4). The format introduced here presents a similar structure than physical RDF data stream formats used in stream reasoning [11] which are defined as a pair SRDF = [(s, p, o), [tS , tE )] where [tS , tE ) are the set of right open time intervals, being (s, p, o) the subject, property and object defined in the RDF format [12]. This situation representation uses a similar approach to [13], defining situations like collections of pairs

Fig. 4. Situation representation format

Parameter-Value in the sub-class scope following the scope idea in [14]. The subclass Type is used as interlink with external descriptors like ontologies defining semantic sensors characteristics [15]. Figure 4 shows the model (top) and its instantiation for the Rut scenario. The Situation Composer Agent sends to the Reasoner Agent, situations which are instant atomic knowledge representations of the world, used as building blocks to formulate arguments. A data flow schema is presented in Figure 5, where the sup-processes used to build situations is represented. Let us highlight that Figure 5 shows the implementation for the scenario introduced in Section 2, focused in sleep activity tracking, but we believe that this approach could fit in other scenarios. 3.4

Justifying and Explaining Behavior

One of the key aspects of ALI, is the capacity of explaining and justifying activities of individuals through the Reasoner Agent. This agent performs several tasks: – Obtain an ordered set of goals, establishing a priority or preference about which goal-based activities that an individual prefers.

XML format

HTTP Rest composer

Sleep Tracking

Social Tracking

...

Other

Transactions Agent

rules

Situation Composer Agent

Situation Composer

ANDROID JAVA SDK

raw data

Relevant Data Selection (highest - lowest values)

Accelerometer Data

Gyroscope Data

Microphone Data

Android SDK ANDROID PLATFORM

SENSORS

Sampling Sensor Agent

Android Mobile Phone

Accelerometer

Gyroscope

Microphone

Fig. 5. ALI-MOB sensor data detection and situation composition

– Build a symbolic knowledge base using observation rules, which are logical representations for the situations captured by sensors. – Build arguments using predefined goals and the knowledge bases. At this point ALI uses all the information related to the defined goal, obtaining relevant data from ACKTUS and building a set of arguments. We use a possibilistic argumentation-based decision making framework [16] in order to deal with uncertain and incomplete data with no data training, inferring more than one possible scenario in polynomial time. – Obtain “strong” arguments using goal preferences. This decision making framework defines some criteria (a semantic argumentation for a possibilistic argumentation framework) for selecting “the best fitted” arguments. We illustrate the process of recognize and explain human behavior performed by ALI in Figure 6. In this figure, only the main Agents are depicted, avoiding those which performs communications tasks for clarifying purposes. 3.5

Recommendation and Positive Feedback

The generation and provision of recommendation and feedback messages is performed by the Recommender Agent, using preferred arguments from the Reasoner Agent. The recommendations are notifications in the form of messages, which are sent back to Rut’s mobile (Figure 3). The notifications are triggered by goals

Situation Agent

Reasoner Agent

1 3

Coach Agent Recommender Agent

4 2 Fig. 6. Justification and Recommendation Process in ALI

and situations, implying its real time nature. In Figure 6 it is represented a 4 main steps flow diagram in order to illustrate the recommendation process in the current scenario. The recommendations are i-nodes from ACKTUS, which contains suggestions about how to modify behavior taking into account detected situations. For instance, the Recommender Agent has defined some test criteria to evaluate sleep using data from [17–19]. When data detected from situations are out of these parameters, a notification alert is triggered like in the case that Rut lies down in bed, and movement is detected after some pre-defined time-period, the recommendation sent to Rut is: “If you are unable to fall asleep within 15 to 20 minutes, go enjoy your sofa and see if there are interesting news to read ”[20]. The test criteria data were:nigh-time: from 21:00 to 5:00, night time insomnia: sleep less than 6,4 hours / night, daytime sleepiness: more than 3,1 hours / day, early awakening: awake at 4:00 and average daily sleep time: 9,5 ± 2,0 (mean ± standard deviation) hours. The idea behind recommendations messages is encouraging the individual to perform activities, in which he/she wants to be supported in order to have behavior change. These messages are formulated by the individual him or herself, supervised by the health care team and stored in ACKTUS repositories. In order to evaluate a real behavior change, a time-stamp is added to the recommendation and stored by the Tracker Agent with the set of situations. The therapist and Rut will be able to see if Rut followed the recommendations and if there has been a change of behavior pattern over a period of time.

4

Related Work

Positive results have been reported from different areas using behavior change support systems, defined as “information systems designed to form, alter or rein-

force attitudes, behaviors or an act of complying without using deception, coercion or inducements” [21]. Randomized trials using Internet based software with positive results in different areas are compiled in [22, 23, 21]. An Internet-based behavior change system using a Bluetooth connected wrist-worn accelerometer to measure physical activity and provide feedback to participants is presented in [24], this work presents a behavior change approach for promote physical activity. Our approach differs from Hurling’s system in giving recommendations promoting behavior change in real time, with no devices attached to individual’s body, which was a drawback in that work, and that is identified as an inconvenient design approach for different authors [25, 26] in the behavior change area. Different approaches using embedded mobile sensors have been developed in order to monitor sleep disorders [27–29] but none of them offer individuals support for behavior change.

5

Conclusions and Future Work

In this work we introduce ALI, an argumentation-based multi-agent system, supporting behavior change in individuals with social withdrawal and mild depression. This work was exemplified by Rut, a mildly depressed woman with cognitive impairments and sleeping disorders feeling socially isolated. The purpose of ALI was to empower the individual, offering persuasive recommendations to change her habits and stimulating those behavior changes through of positive feedback for her actions. ALI is based on the capacity of explaining and justifying activities of individuals using a knowledge repository, in our case ACKTUS repositories. We use a possibilistic argumentation-based decision making framework in order to obtain argumentative explanations for the individual behavior. This framework allows dealing with uncertain and incomplete observations from the individual’s context, with no data training, inferring multiple scenarios in polynomial time. The contributions of this work are summarized as follows: – ALI, a modular agent-based ambient assisted living system for supporting behavior change in mildly depressed and socially isolated individuals. – A supervised platform for monitoring behavior change, applied and exemplified for sleeping disorders and implemented for social interaction analysis. – A situation representation format. A RDF/XML-style format for encapsulate situations, combining stream data systems and traditional knowledge representation approaches. – A model for capturing evidence about behavior change in mildly depressed and socially isolated individuals in home environments, integrating ACKTUS knowledge repositories and applications [7]. Future work includes methods for integrating more sensors in an AAL environment. Different knowledge sources will be integrated for extending and improving the person-specific knowledge based on arguments. A pilot study will be conducted within the near future with four individuals, focusing sleep patterns

and recommendation provision, for the purpose to gain initial results on user experience, usability and relevance for the intended purpose. Moreover, more extensive validation studies are needed with potential users over a time period.

Acknowledgment This research has been partially supported by VINNOVA (The Swedish Governmental Agency for Innovation Systems) and the Swedish Brainpower.

References 1. N. D. Lane, Y. Xu, H. Lu, A. T. Campbell, T. Choudhury, and S. B. Eisenman, “Exploiting social networks for large-scale human behavior modeling,” Pervasive Computing, IEEE, vol. 10, no. 4, pp. 45–53, 2011. 2. J. Yang, “Toward physical activity diary: motion recognition using simple acceleration features with mobile phones,” in Proceedings of the 1st international workshop on Interactive multimedia for consumer electronics, pp. 1–10, ACM, 2009. 3. V. Villarreal, J. Bravo, and R. Herv´ as, “Towards ubiquitous mobile monitoring for health-care and ambient assisted living,” in International Workshop on Ambient Assisted Living, vol. 2010, 2010. 4. H. Sun, V. Florio, and C. Blondia, “A Design Tool to Reason about Ambient Assisted Living Systems,” ISDA, vol. 2, pp. 1136–1144, Oct. 2006. 5. T. Kleinberger, M. Becker, and E. Ras, “Ambient intelligence in assisted living: enable elderly people to handle future interfaces,” Ambient interaction, pp. 103– 112, 2007. 6. M. O’Grady, C. Muldoon, and M. Dragone, “Towards evolutionary ambient assisted living systems,” JAIHC, vol. 1, pp. 15–29, 2010. 7. H. Lindgren, P. Winnberg, and P. Winnberg, “Domain experts tailoring interaction to users–an evaluation study,” INTERACT 2011, pp. 644–661, 2011. 8. H. Lindgren, D. Surie, and I. Nilsson, “Agent-Supported Assessment for Adaptive and Personalized Ambient Assisted Living,” in Trends in Practical Applications of Agents and Multiagent Systems (J. Corchado, J. P´erez, K. Hallenborg, P. Golinska, and R. Corchuelo, eds.), vol. 90 of Advances in Intelligent and Soft Computing, pp. 25–32, Springer Berlin / Heidelberg, 2011. 9. C. Ches˜ nevar, J. McGinnis, S. Modgil, I. Rahwan, C. Reed, G. Simari, M. South, G. Vreeswijk, and S. Willmott, “Towards an argument interchange format,” Knowledge Engineering Review, vol. 21, no. 4, pp. 293–316, 2006. 10. Android, “API Guides - Motion Sensors,” 2013. 11. A. Bolles, M. Grawunder, and J. Jacobi, “Streaming sparql-extending sparql to process data streams,” The Semantic Web: Research and Applications, pp. 448– 462, 2008. 12. O. Lassila, R. R. Swick, et al., “Resource description framework (rdf) model and syntax specification,” 1998. 13. X. Su, E. Gilman, P. Kwiatkowski, T. Latkowski, A. Pr¨ obstl, B. W´ ojtowicz, and J. Riekki, “Knowledge-based systems for ambient social interactions,” Ambient Intelligence, pp. 61–71, 2011. 14. M. Knappmeyer and e. a. Kiani, Saad Liaquat, “Contextml: a light-weight context representation and context management schema,” in ISWPC 2010, pp. 367–372, IEEE, 2010.

15. H. Neuhaus and M. Compton, “The semantic sensor network ontology,” in AGILE Workshop on Challenges in Geospatial Data Harmonisation, pp. 1–33, 2009. 16. J. C. Nieves and R. Confalonieri, “A Possibilistic Argumentation Decision Making Framework with Default Reasoning,” in Fundamenta Informaticae 113, vol. 113, pp. 41–61, 2011. 17. F. M. Xavier and e. a. Ferraza, Marcos, “The dsm-iv minor depressiondisorder in the oldest-old: prevalence rate, sleep patterns, memory function and quality of life in elderly people of italian descent in southern brazil,” International journal of geriatric psychiatry, vol. 17, no. 2, pp. 107–116, 2002. 18. M. A. Carskadon, E. D. Brown, and W. C. Dement, “Sleep fragmentation in the elderly: relationship to daytime sleep tendency,” Neurobiology of aging, vol. 3, no. 4, pp. 321–327, 1983. 19. J. F. Desforges, P. N. Prinz, and e. a. Vitiello, “Sleep disorders and aging,” New England Journal of Medicine, vol. 323, no. 8, pp. 520–526, 1990. 20. M. CM, C. C, S. J, S. R, and B. D, “Behavioral and pharmacological therapies for late-life insomnia: A randomized controlled trial,” JAMA, vol. 281, no. 11, pp. 991–999, 1999. 21. H. Oinas-Kukkonen, “Behavior change support systems: A research model and agenda,” Persuasive Technology, pp. 4–14, 2010. 22. V. Strecher, “Internet methods for delivering behavioral and health-related interventions (ehealth),” Annu. Rev. Clin. Psychol., vol. 3, pp. 53–76, 2007. 23. H. Oinas-Kukkonen, “Behavior Change Support Systems: The Next Frontier for Web Science,” in WebSci10, (Raleigh, NC: US), 2010. 24. R. Hurling, M. Catt, and e. a. Boni, “Using internet and mobile phone technology to deliver an automated physical activity program: randomized controlled trial,” J Med Internet Res, vol. 9, no. 2, p. e7, 2007. 25. G. L. Kreps and L. Neuhauser, “New directions in ehealth communication: Opportunities and challenges,” Patient Education and Counseling, vol. 78, no. 3, pp. 329– 336, 2010. 26. S. Consolvo, D. W. McDonald, and J. A. Landay, “Theory-driven design strategies for technologies that support behavior change in everyday life,” in 27th international conference on Human factors in computing systems, pp. 405–414, ACM, 2009. 27. N. Oliver, F. Flores-Mangas, and R. de Oliveira, “Towards wearable physiological monitoring on a mobile phone,” Mobile Health Solutions for Biomedical Applications, pp. 208–245, 2009. 28. N. Oliver and F. Flores-Mangas, “Healthgear: a real-time wearable system for monitoring and analyzing physiological signals,” in Body Sensor Networks 2006, pp. 4–pp, IEEE, 2006. 29. A. Go˜ ni, A. Burgos, and e. a. Dranca, “Architecture, cost-model and customization of real-time monitoring systems based on mobile biological sensor data-streams,” Computer methods and programs in biomedicine, vol. 96, no. 2, pp. 141–157, 2009.

Department of Computing Science Umeå universitet, 901 87 Umeå www.cs.umu.se

ISSN 0348-0542 ISBN 978-91-7601-136-2 UMINF 14.20

102