Socially Emotional: Using Emotions to Ground Social Interaction

Dolores CafiameroVrije Universiteit Brussel (\rUB)Artificial Intelligence Laboratory

Pleinlaan 2, 1050 Brussels, Belgiumlola@arti.vub.ac.be

Walter Van de VeldeRiverland Next Generation

Research GroupExcelsiorlaan 42, 1930 Zaventem, Belgium

wvdv@riv.be

Abstract

In this paper we address the problem of how emo-tions ground perception and social interaction. Wepropose a virtual society in which each agent has itssubjective perception of the others, and in which thisperception is aimed at focusing a large potential of in-teraction into the most productive direction. Ratherthan to focus attention based on a cognitive evaluationof compatibility between different agent’s interests, weargue for using features of appearance, presentationand material exchange as elements in a "social fil-ter" that directs attention in an emotional as opposedto a cognitive (knowledge-based) way. ~,Ve show howsuch external clues are tight into a hormonal modelof emotions for behavior control. At the social levelwe develop ideas on how emotions play a role in agentinteractions (emotional rhetoric) and long-term agentcoordination (representation).

Introduction

The problem we address is the following: how do agentsthat are embedded in large societies--hundreds to mil-lions of agents--focus the enormous potential for in-teraction with others towards the most productive en-counters?

The relevance of this problem needs little argu-ment. From a social studies viewpoint, it is tied tothe understanding of a society’s structure and howit--implicitly and explicitly, directly and through in-ternalization mechanisms--regulates daily interaction.From a multi-agent (MA) studies perspective, the is-sue is fundamental to understanding artificial societies.In particular, it enlarges the scope of behavior controlin agents, which is not only aimed at task achieve-ment (e.g., finding food to survive), but at doing so an evolving social context that, by the interdependen-cies it entails, makes every strict assumption of self-sufficiency a naive one. More technically, one cannotexpect exhaustive exploration of potential encountersto be an efficient means of implementing large-scaleagent societies. This is especially true for open MA

systems--those in which the agents are being designedindependently and without a predefined and globallyknown relationship towards the others. Finally, andhere lies an important motivation for our work, suchlarge-scale open agent societies will soon be of practicalrelevance---we assume that networks like WWW willevolve into large collections of software agents, eachindependently pursuing goals on behalf of their users,or trying to offer services to users and other agents.

This paper is concerned with the development ofconcepts and techniques that can enhance a variety ofsocial processes in large, informally structured groupsof agents inhabiting different types of realities. Thekind of applications this research is aimed at involveinformation-intensive events in which the possibilitiesfor interaction among participants are many, but notevery encounter leads to a productive social process.For example, at a major fair there may be many po-tential buyers and sellers, but only some will come to aproductive interaction--a business deal. Our practicalgoal is to focus this large potential for interaction sothat the effectiveness of participation to such events isenhanced. To foster this goal, our work brings togethertwo lines of research----one focusing on social dynamicsin mixed-reality agent societies, the other addressingthe problem of achieving flexible and adapted beha~,-ior in (societies of) self-sufficient, autonomous agents.

Co-Habited Mixed Realities

The particular context of our work is set by the con-

cept of co-habited mixed reality (\ran de Velde 1997),developed within the COMRIS research project 1. Aco-habited mixed reality consists of a pair of a realand a virtual space that are loosely coupled--there is aminimal awareness of one space within the other. Thereal world is inhabited by humans, whereas the virtual

~Project COMRIS (LTP~ 25500, 97-2000), approved forfunding within the EU Long-Term Research initiative Intel-ligent Information Interfaces (I3). The views in this paperare not necessarily those of the COMRIS consortium.

10

From: AAAI Technical Report FS-97-02. Compilation copyright © 1997, AAAI (www.aaai.org). All rights reserved.

one is inhabited by software agents. The originality ofthis concept relies on the type of relationship betweenboth realities, which also gives it its strength. Contraryto augmented-reality models, we are relieved from thetask of establishing a mapping between both realities,as well as from having to try any kind of perceptualintegration. Instead, the processes in the two spacesmust only serve each other, without the real partici-pants having a feeling of virtual presence or the otherway around. We only provide a personal link betweenboth spaces, that ensures some form of "synchroniza-tion" so that the activities in one space further theones in the other. This link is an electronic badge andearphone device integrating a vision system, wirelesslyhooked into an Intranet, which each real participantwears as her personal assistant--the COMRIS parrot.It establishes a bidirectional link between both reali-ties. It perceives what is going on around its host (con-text perception) and points her to relevant informationso that participation in the ongoing event is enhanced(information push). At the same time, a similar eventis going on in the virtual space, where software agentsare the participants. The purpose of each agent is torepresent, defend, and further a particular interest orobjective of the real participant, including those she isnot explicitly attending to.

To achieve this, the agents must be autonomous andself-motivated. Techniques of interest-based navigationbring together those virtual agents whose interests arelikely to merge into a productive social process. Todeal with the complexity of finding out about usefulpartners we propose the notions of appearance--the ex-ternal "bodily features" that partly reflect the agent’sinternal state--and of presentation--the set of featuresthat the agent shows to the agents it interacts with, de-pending on its interests in the communication process,its expectations about the others, etc. The differentrepresentatives of a user may have to compete to gettheir owner’s attention at a given point. At all times,competition ]or attention mechanisms (Van de Veldeet a1.1997) that take into account several factors con-tributing to define the internal state of the agent, suchas its competence, its performance, or the relevance ofits action in a particular information context, drive thecommunication with the real participant through theCOMRIS parrot.

In dealing with large-scale interaction potentials inthe COMRIS context, two different types of scenariosare of particular relevance. A "synchronous" scenarioinvolves the interaction between the software agentsand a user by means of the parrot. Our main contri-bution here lies in linking a bodily-grounded emotion-based behavior control framework with the relevance-

and competence-based attention mechanism.The other scenario has to do more with the inter-

action patterns that are going on within the virtualworld, regardless of the link with the real world. Inaddition to the above-mentioned mechanism of compe-tition for attention, we are specifically focusing on thelonger term, "asynchronous" coordination of agents’behaviors. If, as we argue, social interactions arebased on constellations of agents’ interests that can befurthered by those interactions, then long-term socialstructures must rely on some explicit or implicit encod-ing of these interest relationships. Instead of a purely"cognitive" encoding in which the agent accumulates amemory of other agents’ (perceived or communicated)interests, we propose to explore alternatives that areno doubt being used in real societies. For example,values and emotional states play a major role as "fil-ters" in learning and remembering social interactions.Appearance acts as another filter by expressing powerrelationships, clan or sub-culture membership, status,and so on. Similarly, the ritual of presentation2 al-lows one to express one’s intentions and expectationswith respect to a particular party in a particular situa-tion. Finally, material goods that are being exchangedrepresent such intentions and expectations over longerperiods of time, thus regulating and coordinating be-haviors of two or more agents.

The concrete direction that we explore here is to usebodily-grounded emotions and their various expressionsas an essential part of appearance, presentation, andthe emergence of long-term co-behaviors. Our softwareagents are also endowed with other affective phenom-ena, equally rooted in their physiological state, such as"survival-related" motivations driving them to act au-tonomously, moods, and temperaments that contributeto define their individual personalities.

Why Emotions?The nature and role of emotions in biological systemsare still not completely agreed upon nowadays. Severalcomponents are usually distinguished in the develop-ment and expression of emotions (Kandel et al.1995):the recognition of an important event; an emotionalexperience in the cortex that mediates outgoing sig-nals to peripheral structures; and reflexive autonomicand visceral responses that prepare the body for ac-tion, some of which are also externally observable. Dif-ferent aspects and the causality direction of the linksamong them have been stressed by diverse research

2The behavior that precedes a purposeful act together,and that can be as simple as exchanging a business cardor as complex as marriage negotiations in certain societies(Keane 1997).

11

paradigms (see (Cornelius 1996) for an overview). multi-facetedness of these phenomena, which encom-pass neuro-endocrine, physiological, cognitive, and so-cial aspects of animal behavior, make them a very richtool to ground adaptation, autonomy, and social inter-action in our society. Some of the roles that emotionalstates seem to play in biological systems are of partic-ular significance for our work.

Emotional states are adaptive mechanisms that al-low to deal with events, both internal and external,which are important for the survival of the agent.(and the species) in a particular environment (LeDoux1996). In this sense, they are considered as remnantsof the creature’s evolutionary past. The perception ofa potentially life-threatening event, such as the pres-ence of a predator, immediately puts the animal in a"fear state"; this state acts as a defense mechanism byeliciting a "fight-or-flight" reaction that prepares theorganism to respond rapidly to the external environ-ment (for example, a transient, rapid increase in bloodpressure permits the animal to run faster).

Emotions play a major role in motivating and guid-ing action. At the simplest level, the categorizationof events as pleasant/unpleasant, beneficial/noxious,turns "neutral" stimuli into something to be pursuedor avoided. This idea lies at the heart of operant con-ditioning theories of learning. Due to their generalityof object and time, emotional states amplify the ef-fects of motivations, which are stimulus-specific andneed an urgent satisfaction. As (Tomkins 1984) pointsout, "the affect system is the primary motivationalsystem because without its amplification, nothing elsematters, and with its amplification, anything else canmatter." Tim role that emotional systems play in de-cision making (activity selection) and social relationshas been evidenced by Damasio’s studies of patientswith damage in the prefrontal cortex and the amyg-dala (Damasio 1994).

Finally, the external manifestations of emotions canplay a major role as "signaling" mechanisms3, at sev-eral levels. First, they play a role of social reference(Dantzer 1994). The emotional expression of an indi-vidual can be used by another to "assess" the type ofsituation it is confronted with and adapt its behavioraccordingly. Pheromones4 can help to illustrate thispoint. For example, the smell of the urine of a rat ex-posed to electric shocks induces in another rat an in-crease of the pain sensitivity threshold. In some cases,

3This does not imply that emotions evolved primarilyas a communication tool.

4Chemicals released in the environment by the organismthat influence the physiology and behavior of conspecificsand sometimes individuals of other species.

emotional expression seems to have a communicativerole, as in the case of the diverse alarm calls emittedby monkeys in different fear-related situations and thatelicit different behaviors in conspecifics (Kalin 1997);alarm calls (and emotional expression in general) canalso vary depending on the "receptor", as reported in(Dantzer 1994). Emotional states also contribute the construction of intersubjectivity. The recognitionof the emotional state of a conspecific allows an ani-mal to have certain expectations with respect to theother’s behavior, but also to adapt its behavior andemotional state accordingly. For example, an animalwill rarely attack another that shows signs of submis-sion/sadness. In this sense, emotions contribute to theelaboration of the image an individual has of itself withrespect to the partner it is interacting with. This self-image is a reflection of the image we have of the other,and elicits the mirror process in it, which in turn canelicit an emotional/behavioral reaction on my side thatshows my agreement of the image the other has con-structed about me, gi~4ng thus rise to the constructionof a negotiated intersubjectivity.

Bodily-Grounded Emotions

Our model of motivations and emotions is based on(Cafiamero 1997a), which draws on ideas from neuro-biology, ethology, behavior-oriented AI, and artificiallife to explore the effects of affective phenomena in be-havior. This work stresses the role of affects as mech-anisms for adaptation, both from an evolutionary per-spective and in the sense that this term has in (Ashby1952)--contributing to maintain the organism (a dy-namic system) viable in its environment, i.e., internally"stable". Therefore, emotional states are not directlymodeled at a "psychological level", but grounded in aphysiology that ensures appropriate feedback betweenbehavior and affective states in the interactions of theagent with the world. The original model, describedin the next section, considered emotional states at theindividual level. Here we propose an extension to dealwith "social emotions" within the same framework.

Emotions in Viability and Adaptation

The experimental setting to explore these ideas in(Cafiamero 1997a) is a dynamic two-dimensional worldinhabited by two species--the "smart" agents (As)and the "enemies" (Es)--and which also contains foodand water sources and blocks--geometrical figures--of varying shapes and sizes. The creatures are im-plemented as a collection of "modules" of differenttypes. Es have a simple, rather stereotyped behav-ior, their main role being to introduce more dynamic-ity in the world--they wander around avoiding obsta-

12

cles, try to ingest every form of organic matter theyrun into, and they withdraw whenever they feel pain,caused by another creature’s attack. As are more com-plex creatures used as a testbed for affects. Amongother elements, their architecture includes: internaland external sensors~ maps, behaviors, both appeti-tive (go-toward, look-for, find) and consumatory be-haviors (eat, drink, play, rest, withdraw, etc.); motiva-tions (e.g. fatigue, thirst, boredom) and basic emotions(fear, anger, happiness, etc.). In addition, they have synthetic physiology--a set of parameters that definetheir internal bodily state and needs. This physiol-ogy includes both, controlled variables necessary forsurvival (e.g. heart-rate, energy, blood sugar level)and hormones released under different emotional statesthat modify the amount of the controlled variables.Controlled variables are monitored by motivations, im-plemented following a homeostatic model. Each moti-vation receiving an error signal (whenever the value ofthe corresponding controlled variable moves away fromits ideal value) gets an activation level proportional totile magnitude of the error, and an intensity calculatedon the basis of its activation level. Several motivationscan be active at the same time, but that with the high-est activation level gets the creature’s attention andtries to organize its behavior so as to satisfy its maindrive. The strongest motivation5 selects the (consuma-tory or appetitive) behavior(s) that can best contributeto the satisfaction of the most urgent need. The exe-cution of the selected behavior modifies the values ofdifferent controlled variables with an intensity propor-tional to that of the motivation, bringing the variablethat triggered the drive to its ideal value, and thereforesatisfying the need that originated the behavior. Moti-vations thus drive behavior selection and organizationbased on the notions of arousal and satiation, and alsodetermine the agent’s focus of attention. They can bethought of as implementing an implicit value systemthat ensures decision making, activity selection, andautonomy. Emotions are "second-order" modifiers oramplifiers of motives and therefore of behavior. Un-like motivations, the homeostatic model do not seemscompletely adequate in this case, in particular the no-tions of arousal and of activation as its indicator. Con-trary to activation theories, we follow (Pribram 1984)to view emotional activation as an indicator of a changein configuration of neural and endocrine activity withrespect to the habitual stable baseline of the organism.Emotional states are thus signs of internal "instabil-ity", but they also seem to play a role in homeostasisas mechanisms of re-equilibration, which derived from

5A second motivation is also taken into account to allowfor oportunistic behaviors.

processes that try to stop ongoing behavior, comple-menting this way the role of motivations--appetitiveprocesses that try to activate action as a response todeprivation. In our model, emotions are implementedfollowing a hormonal model (Cafiamero 1997b). Theyexert further control of the agents behavior by send-ing synthetic hormones that may affect not only theagent’s controlled variables, but also its perceptual.attentional, and motivational-mechanisms, this waymodi~;ing the intensity and execution of the selectedbehavior. The inclusion of emotions in autonomousagents fosters flexibility of behavior and adaptation.

The integration of the original model in the COM-RIS agents requires some modifications and extensionsto accommodate the demands of complex social inter-actions, as well as those of an information context. Letus consider the main ones. First, the physiology ofthe agents must be augmented to include parameterswhich are relevant for an agent that has to survivein an information context: parameters to keep trackof the amount and relevance of information, of theagent’s performance, of its competence, etc. Motiva-tions, implemented again according to a homeostaticmodel, monitor these variables by selecting behaviorswhose execution modifies the parameter in the desireddirection. In the same way, the particular modules in-tegrating the agents (sensors, maps, behaviors, etc.)must be adapted to this particular world. As for emo-tions, we propose to complement the set of basic emo-tions used in the original model, each implementing anadaptive function fundamental for the survival of anagent in its particular environment, with a set of "so-ciai emotions" that come to play only in the context ofsocial interactions, and that should contribute to theadaptation of the agents to their social environment.Therefore, the main extensions have then to do withthe ability to show, perceive, and react to emotionalexpression, as we will see in the next section.

Negotiating Difference: EmotionalRhetoric

The model of "social emotions" that we propose hereis inspired by that of Aristotle in his Rhetoric, whichconsiders emotional states as tools to negotiate socialinteractions. This view starts by accepting the factthat individuals are different, and that they ha:ce to livetogether; in order to do so in harmony, they must takeinto account their differences and accept them or nego-tiate them. Emotions are signs of a difference betweenindividuals, and at the same time the tools that al-low them to negotiate this difference, fulfilling variousroles---epistemic, rhetoric, ethical, and political. Themodel of affects that Aristotle proposes in its Rhetoric

is based on a couple of relationships between individ-uals: identity/distance (according to their interests,affinities) and symmetry/asymmetry (rank, status), perceived by each partner. A social interaction be-tween two individuals will produce an emotional statein each of them according to its own perception of therelation along these dimensions and of the image thatthe other shows about it. The effect of an emotionalstate~ as far as social interaction is concerned, will beto try to maintain; increase or decrease the distance ineach dimension, through external manifestation suchas emotional expression and/or the execution of anappropriate behavior. For example, love between Aand B implies the nmtual perception of an "almost-identity" (empathy) and of symmetry that both in-dividuals will try to maintain or increase. Emotionscan also activate each other and form "rhetoric" chainsto negotiate distance and symmetry. For example, anoverconfident individual A engaged in an interactionwith B will perceive itself as having a much higherrank than B, disregard the "distance" dimension, anddisdain whatever B proposes or does towards A. If theperception that B has of this relation is that A overes-timates its self-image and underestimates B (i.e., therelation is not as asymmetrical as A seems to believe),B will enter in an angry state that will lead it to dosomething that increases the distance and decreasesthe asymmetry as perceived by A ("Hey, who do youthink you are talking to?!!"). Emotions form a sortof interactive network, with opposition (e.g., love ver-sus hate) and inhibition relations (e.g., anger and fearcannot be active at the same time) as well.

In the "original" COMRIS model the agent featuresof appearance, presentation, competence and relevanceare used to regulate interest based navigation and com-petition for attention. Having adapted the agent’semotional apparatus, we can take the crucial step of in-volving emotional expression in each of these factors.Emotional expression can be kept simple, and how-ever rich enough to meet the requirements of both ap-pearance and presentation. Each emotional state auto-matically carries some external features over which theagent has no control (for example, a color associatedwith each emotion) and that affect its appearance~apattern that reflects the agent’s self-perceived "rank"or competence regardless of other individuals. This"rank" or competence level can be a function, for in-stance, of the history of successful interactions (e.g.,those that increased the "pleasure" variable above acertain threshold), and the amount and relevance ofaccumulated information or wealth. Ranks can belongto a culturally determined typology, and establish hier-archical relations in the society. The "rank" of an agent

then tells at least as much about its social "type" (towhich a social role can be associated) as it does aboutits history (how successful it has been in its interac-tions); it is thus a "social identifier". In addition, theagents should be able to control some of the compo-nents of emotional expression so that they can choosehow to present themselves to the others depending ontheir state, interests, and the image the have of theother, i.e., of their view of the distance and symme-try dimensions of the relation. Presentation can beexpressed by modi~’ing the pattern used to show ap-pearance as a function of the "social" emotional state.In this way, presentation reflects the agent’s self-imagein the context of a social relation, reflecting also theimage it has of its partner. This image of self and theother has also a behavioral impact in the interaction;for example, in the "price" that an agent proposes toexchange goods or information with another. As foremotion recognition, we propose the use of expressivepatterns characterizing the different emotions. In par-ticular, the difference between the appearance and pre-sentation patterns of an agent A can be used by an-other agent B to recognize the image that A has of therelation, and therefore of B. A comparison of this dif-ference (A’.s point of view) with the difference betweenB’s own appearance and recognition patterns (B’s viewof the relation) will trigger a particular emotional statein B that will have an impact on B’s physiology, andproduce an appropriate behavioral reaction towards Aintended to negotiate the difference between their im-ages. The recognition process does not need to be cog-nitive itself. For instance, a mechanism similar to theuse of pheromones for rank and emotion recognition insome mammals can serve to implement the recognitionof appearance and presentation, and the effects thattheir difference produces in the agents’ physiology andbehavior. In cases when it is important to maintaina history of individual interactions with other agents(or of interactions with types of agents), this differencein viewpoints can be kept as an emotional memory or"snapshot" that bootstraps the process in future inter-actions.

Conclusion and Prospect

The main roles that emotions play in social interac-tions in COMRIS can be summarized according to thescenarios that we distinguished previously.

At the individual agent level, emotions ground thenotions of competence and relevance. Emotional statesbeing in part a consequence of the agent’s competenceand relevance, as reflected in the level of the pertinentphysiological parameters, these two notions are mani-fested in terms of emotional expression. This way, they

14

have a direct impact ill social interaction. In turn, thelevel of success in this interaction feeds back into theagent’s emotional state.

In the synchronous scenario, emotional states playa major role in the agents competition for the user’sattention. Similarly, emotional expression is used bythe parrot to solve the competition for attention game.Ideally (but admittedly not easy) this should imply theselection of agents that seem to best fit not only thecurrent interests of the user (e.g., the agents that lookmore self-confident) but also her emotional state (e.g.,the parrot may decide that a sad user might need ahappy agent).

In the asynchronous scenario, we envisage two mainroles for emotions. One is the negotiation of differencebetween individuals--of their "proximity" in terms ofinterests and of the "symmetry" of their relation asperceived by each partner. An interesting issue to ex-plore here is whether, and under what circumstances,the local interactions between couples (or small groups)of agents can lead to the equilibrium of the society, asclaimed by Aristotle for his City, i.e., the political roleof emotions. The other role is in the creation of long-term interest relationships. For example, a very self-confident agent unable to achieve a goal on its own mayhave to call for another agent’s help. This can make thefirst agent to feel "ashamed" because it was not self-sufficient, and create a debt towards the second agentthat it might want to pay back only privately--withoutthe presence of witnesses that would know about itsweakness. Or a sad agent that received the gracioushelp of another one can develop a long-lasting grati-tude debt towards it, possibly implying more than oneaction at different times to show (and pay back) thisgratitude. A concrete case that we have explored inmore detail--and which is also well studied in anthro-pological literature--is the ritual act of present giving(Keane 1997). A present, whatever physical object is, becomes a re-presentation of the present-giving act,which is embedded in a complex socially determinedscript of recoguition and obligation. By its persistentnature and the emotional binding that both giver andreceiver have with it, it creates co-behavior patternsthat can extend over longer periods of time, especiallywhen the exchange was witnessed by others (think ofmarriage rituals). We are proposing a scheme in whichagents can exchange "material" tokens, that they as-sociate with an "emotional snapshot" of the exchangeevent. In later dealings with the other agents involved,the token-snapshot combination triggers a (culturallydetermined) emotional pattern. In this way, then, wehope to bridge the gap between bodily-grounded emo-tions and their social roles.

AcknowledgementsDolores Cafiamero is supported by the Spanish Min-istry of Education under grant PF95-00410164. Wal-ter Van de Velde is an honorary senior researcher forthe Flemish Science Foundation (FWO).

ReferencesAshby, W.R. 1952. Design ]or a Brain. London:Chapman and Hall.

Cafiamero, D. 1997a. Modeling Motivations and Emo-tions as a Basis for Intelligent Behavior. In W. LewisJohnson, ed., Proceedings o[ the First InternationalCon]erence on Autonomous Agents, 148-155. NewYork, NY: ACM Press.

Cafiamero, D. 1997b. A Hormonal Model of Emo-tions for Behavior Control. VUB AI-Lab Memo 97-06, Vrije Universiteit Brussel, Belgium.

Cornelius, R.H. 1996. The Science o[ Emotion. UpperSaddle River, N J: Simon & Schuster.

Damasio, A.R. 1994. Descartes’ Error. New York,NY: G.P. Putnam’s Sons.

Dantzer, R. 1994 (2nd ed). Les dmotions. Paris:Presses Universitaires de France.

Kalin, N.H. 1997. The Neurobiology of Fear. ScientificAmerican Special Issue 7 (1): 76-83.

Kandel, E.R., Schwartz, J.H., Jessell, T.M. 1995. Es-sentials of Neural Science and Behavior. Norwalk,CT: Appleton & Lange.

Keane, W. 1997. Signs of Recognition. Powers andHazards o] Representation in an Indonesian Society.Berkeley, CA: University of California Press.

LeDoux, J. 1996. The Emotional Brain. New York,NY: Simon & Schuster.

Pribram, K,H. 1984. Emotion: A NeurobehavioralAnalysis. In Scherer, K.R. & Ekman, P. Approachesto Emotion, 13-38. Hillsdale, N J: Lawrence Erlbaum.

Tomkins. S.S. 1984. Affect Theory. In Scherer, K.R.& Ekman, P. Approaches to Emotion, 163-195. Hills-dale, N J: Lawrence Erlbaum Associates.

Van de Velde, W. 1997. Co-Habited Mixed Realities.Proceedings o[ the HCAI’97 Workshop on Social In-teraction and Communityware, Nagoya, Japan, Au-gust 1997.

Van de Velde, W., Geldof, S. and Schrooten, R. 1997.Competition for Attention. Proceedings o[ the 4th In-ternational Workshop on Agent Theories Architec-tures and Languages, 282-296. Providence, Rhode Is-land, USA, July, 1997.

15