Socially Intelligent Agents

AAAI Fall Symposium Series 1997, held in the Tang Center at MIT in Cambridge, Massachusetts, November 8-10, 1997


The SIA Symposium is sponsored by

Sony

.

Links:



The symposium will discuss the issue of socially intelligent agents (SIA), focusing on the concrete realization of an artificial system (robot, software agent). Discussions consider humans, software and hardware agents, in both natural and synthetic environments. The discussions outline cross-technological concepts (excluding those restricted to a specific hardware or software technology). The first day of the symposium will present SIA research in different communities. The talks address the following issues: modeling of human societies, social robots, believable software agents and software pets, cultural aspects of social intelligence, VR as environments to further human social networking. A game is planned for the second day of the symposium. It will be an opportunity for participants to study strategies of social interaction "from within", i.e. by being part of the social interaction game.

Chair

Kerstin Dautenhahn
The University of Reading
Department of Cybernetics
Whiteknights, PO Box 225
Reading, RG6 6AY
United Kingdom
Fax: +44 (0) 118-9318220
Tel: +44 (0) 118-9318219,
+44 (0) 118-9316372
E-mail: K.Dautenhahn@cyber.reading.ac.uk

Organizing Committee

Michel Aube (Universite de Sherbrooke, Canada)
Joseph Bates (Carnegie Mellon University, USA)
Kerstin Dautenhahn (University of Reading, UK, chair)
Philippe Gaussier (ENSEA, France)
Judith Masthoff (Institute for Gerontechnology, The Netherlands, co-chair)
Chisato Numaoka (Sony Computer Science Laboratory - Paris, France, co-chair)
Aaron Sloman (University of Birmingham, UK)


You can find out more about:

Information for participants:

Please read the following information carefully. I tried to put all relevant information concerning participation and publishing on this Web page.
 In July 1997 AAAI will mail the registration brochure to invitees.

Information for authors:

Information for Speakers:


Objectives of the Symposium

 The symposium will analyze various forms of social interaction, their functions and the preconditions that make them possible, in both natural and synthetic agents, and in physical and software environments. This includes studying the different types of social expertise and their causes and consequences, along with means of checking which agents have them.

The term "social" has become fashionable recently, describing various kinds of interactions between "agents", comprising artificial hardware and software agents as well as animals and humans. Moreover, work on "social dynamics" often combines approaches on different levels of abstraction and involves different degrees of behavioral or cognitive complexity of the agents which are studied. In robotics and multi-agent-systems the term "social" is often used in a sociobiological interpretation, based on game-theoretical concepts. Research on intelligent software agents and artificial characters focuses more on aspects of intelligence like personality, believability or emotions, which some people regard as non-rational.

In all these different fields and applications the term "social" is most often used with a very general, common sense meaning, without providing a basis for the evaluation of "social expertise".

This symposium aims to both enlarge and concretize the issue of social expertise which has been discussed so far in different research areas like psychology, sociology, biology, artificial intelligence and robotics. In particular, we wish to address the orgins and development of social expertise with respect to the concrete realization of an artificial system. This includes both the external behavior and the internal cognitive and motivational abilities of an agent. The concepts discussed should include both software and hardware artifacts. This means that the discussions should focus on cross-technological concepts (excluding those restricted to a specific hardware or software technology). This general focus is necessary in order to find a common language between participants from different fields.

We propose the following assumptions concerning social expertise:

a) Social expertise in natural systems is normally linked to an embodied agent situated in a concrete dynamic environment. The complexity of human social behavior in such contexts is correlated to the complexity of a human body and our ability to perceive and produce subtle bodily changes, including facial expression, posture and tone of voice.

Recently the development of electronic communication has shown that rich social interaction is possible without direct physical contact or mutual physical perception. The speed of electronic communication enables a type of immediacy not previously achievable by paper-based correspondence, and enables new forms of interaction in which physical appearance and even gender of participants have no role since they are not known to others. In future, virtual reality environments may, like masked balls, enrich the forms of social interactions in which people adopt temporary personas linked to temporary physical characteristics.

Despite these variations in context there are common themes:
b) Although artificial agents are still far from the complexity of natural sytems, future robots and software agents will need to interact with each other, and with humans, using types of social expertise that may begin to match human social competence. For some purposes, e.g. in disembodied or distributed agents, new forms of social interaction may be developed.

Several questions for the design of socially intelligent agents arise from these assumptions:

Social Agents as a Technological Challenge

Robotic and software systems which interact with each other and with humans are a challenging technological as well as scientific area of research.

The design and construction of intelligent agents has often been seen as a technological problem (Technology in the sense of constructing the system according to a specification, writing a control program and evaluating the system).

With new physical hardware and engineering expertise we can construct robots which interact in the physical world, with other robots or with humans. These aspects are studied in the field of robotics and in more recent approaches in behavior-oriented robotics and artificial life. In these areas the need to ground any kind of intelligent behavior in sensori-motor skills of a single agent has become a dominant focus of research. Approaches to collective intelligence study how groups of robots can interact and cooperate in a specific environment. Many of these approaches model collective behavior of social insect societies, which is based on local, "anonymous", "reactive", interactions without any individual social relationships, though no existing robots can match the complexity and sophistication of most insect communities.

It is an open question to what extent this research is too narrow: based on restrictive assumptions about the basis of human and animal intelligence, and omitting important aspects of human intellectual competence. One question for the workshop is the extent to which the assumptions need to be generalised, to accommodate a wider variety of architectures for social agents.

In the area of service robotics some approaches study robotic systems which should support humans while having long-term contacts with humans. Such "social robots" in service applications have to exhibit a minimum degree of social expertise which allows them to interact with humans in a way which is "natural" for unexperienced users. Research in this field uses the term "symbiosis" to characterize human-robot relationships. We need to investigate different classes of robot services, including distinguishing those that require human-like physical forms (e.g. surrogate mothers?) from those that don't (e.g. intelligent aids for the blind or disabled).

In the case of software agents ("softbots", "intelligent agents") we can distinguish the following cases.
Artificial agents and humans can also interact with each other in virtual realities. Here, the design of a "social interface" for software agents resembles those problems which designers of "social robots" face, depending on the form of virtual reality and its faithfulness to real physics.

In all cases one of the issues to be addressed is how far the artificial agents and human agents need to share ontologies, values and goals in order to be able to communicate.

Which implementation techniques may suffice in which contexts?

The goal cannot be simply to copy or mimic certain externally perceivable aspects of human social interaction and communication. The complexity (e.g. multi-modality) and variety of human reactions and means of interaction make it unrealistic to capture all potential behavioral reactions of humans in a look-up table which the artificial agents only have to apply properly in real-time. A "shallow" way of modelling could be successful in very restricted areas of application, where the search space of possible social interactions is small and where the purpose of the interaction is narrowly constrained, e.g. in some forms of entertainment. Instead, in real world and very "open" systems, it seems to be necessary to analyze and replicate mechanisms of animal (and human) social interaction and social expertise, for instance the various perceptual, motor, cognitive and motivational modules that coexist within an individual and interact over time. This task involves trying to find a common language for specifying and modelling social behavior which is applicable across technologies and "species" (including humans).

Social Agents and the Origins of Intelligence and Communication

Questions about the nature, origins and mechanisms of social expertise form important topics in recent studies in primatology, developmental psychology and sociology. There is evidence that social interactions are the basis for the development of a concept of person, learning socially and ecologically relevant behaviors. Similarly, social agreement processes are the basis for symbols, mathematical thinking and abstract problem solving. In spite of different opinions on its concrete role and importance, social intelligence is thought to be one of the fundamental factors providing the origins of intelligence in primate evolution. Others claim that more basic forms of intelligence required for interaction with and navigation in the physical environment underlie forms of social competence and learning. At present we need to have an open mind about directions of influence between social and individual competences, until we have a deeper understanding of all the options. Even if there is no straightforward way to transfer knowledge about human behavior to the behavior of artificial agents, the idea seems to be promising. Humans are supposed to be that species where the most complex social interactions can be found. Humans are social animals, this is one of their areas of expertise which can be found across cultures and at every ontogenetical stage. Moreover, neurological studies more and more stress the tight interdependancy between emotions, social expertise and human rational thinking.

Topics

The symposium will focus on studies in human-like social behavior and expertise and on approaches to designing and evaluating artificial systems which interact socially with humans in an acceptable way.

This should cover at least the following topics which should be discussed due to their phenomenology and biological function in human social agents as well as their usefulness and applicability for implementations of artificial social agents:


Up


Program



URL's of the speakers can be found on the Webpage of the Participants.