Agent-based systems, especially those that use lifelike personas in communicating with users, have a special mystique about them, and hold a unique attraction for us. Why? What is it about these embodied personalities that can, on the one hand, be so delightful, informative, and engaging, yet can on the other generate so much controversy? A central theme of this article is that synthetic agents are attractive to us because they communicate in new, often effective ways which have seldom been used in computer applications, and that this communication is primarily social in nature. Here we will look at some of the many research contexts in which the autonomous agents community is building interactive, entertaining, and sometimes uniquely effective software that encourages users to employ their inherent social skills. Some applications are unabashedly for entertainment (such as PF Magic's playful virtual pets), some are unabashedly scientific in nature (such as MIT's BodyChat for manifesting explicit communicative gestures). Yet these efforts, and others in the spectrum they help define, share many common themes. One of these, present in all of the work discussed here, is that people are naturally entertained by many of the emotional, communicative, community-based, attributes of synthetic agents that will also help us to write more effective software, whether agent-based or not.
A savvy thinker might question the wisdom of inflicting a set of non-explicit, dynamic, social rules onto already complex software scenarios. In fact, as might be expected, research has shown personified interfaces to have negatives as well as positives (c.f, [Koda1997], and [Walker & et al1994]). But consider the existing alternative: as it stands now, humans, being much smarter and more flexible than software, have shouldered most of the burden of adapting to the artificial communication protocols established between users and software programs. Preferring speech, we type; preferring conversation, we give commands; designed to negotiate, we get no negotiable feedback. But people are social animals. They have highly developed social skills which are not typically being used to direct, and gather information from, their computers. Progress in reducing the adaptive requirements of human users by increasing the ability of software to understand, and communicate in, the human world is a big win. This is where agent-based software, especially as embodied in agents with redeeming social qualities, is attractive.
New horizons are unfolding as well. In many cases we find users interacting with their computers to help them reach goals not formerly associated with computers. Research is under way to create interesting artificial personalities, to set avatar projections of users into immersive virtual environments, to help users in educational processes through personified interfaces, to develop architectures that support human-computer interaction -- especially so as to increase the usable bandwidth of the information path, and to form models of users' states in the synthetic social fabric.
It is clear that researchers building lifelike agents must guard against misunderstandings, especially in a market so attractive to industry. We personify our agents and label them anthropomorphically because it helps us to communicate with one another and to share our work -- especially so across disciplines. But others not familiar with what to us are the obvious limitations of shallow personified agents may read into our efforts at naturalness an intelligence that does not exist. But is this really just the latest in computer homunculus alchemy? Even if one were to make the dubious claim that too much energy is being spent on creating the illusion of life, the strong counterargument is that developing the necessary AI tools to get this right almost by definition requires that we understand, and can model, human social intelligence - an admirable goal in its own right. To make Extempo's ``virtual tavern'' interesting, one has to understand, and simulate, that which makes bar culture interesting (see below); to have Affective Reasoner agents derive engaging stories to tell, one has to know what makes a story fun to hear (see below). The core issue is in deciding, now that we have platforms to support it, whether or not it is worth investing in building computer systems that can traffic in the same sorts of communications that people have used for thousands of years.
Always there are tradeoffs. Ease and breadth of application are often at odds with theoretical depth, and presentation detail. In some cases, assuring believable personas means picking the easier problems so to avoid inherent pitfalls. For example, among the systems discussed we see PF Magic modeling real creatures with little expressive sophistication in their repertories; in the AR we see a minimalist approach to graphics and timing; with CyberCafe we see a restricted language model. In other cases it means emphasizing theoretical correctness over initial flexibility, such as in the complex cognition and affect subsystem models of Sloman, et al., in the detailed approach to diectic believability of Lester et al., in the finely crafted characters of Bates et al., in the fine-grained gesture analysis of Cassell et al., and in the sophisticated agent emotion models of Elliott, Velazquez, and Botelho. Yet in all cases we find the working out of fluid, plausible, real-time personas requiring great devotion to ``getting it right.'' This helps to form a culture that is learning about the general rhythm of synthetic social intercourse. Hence progress in each of the paradigms here, whether or not successful in the explicitly stated goals, contributes to the real corpus of work necessary to build the next generation of socially interesting, and ultimately entertaining, agents.
In the long run, this is in the best tradition of AI: We can justify the sometimes Herculean tasks set before this research community because in our efforts we hope to mimic, and understand, that which is noble, unique, useful, and robust, about the human social condition itself.