Modeling Physical Personalities for Virtual Agents by Modeling Trait Impressions of the Face: A Neural Network Analysis

[abstract]

The 1990s gave rise to a host of virtual agents: synthetic characters, interface agents, and virtual humans. Although users welcome the prospect of interacting with virtual agents, more often than not these agents disappoint by being too mechanical and inconsistent in their behaviors. In short, they lack personality. Researchers, recognizing the importance of personality in creating socially engaging virtual agents, have sought ways of endowing agents with a convincing artificial personality.

There are many aspects to personality. One pressing concern in this area of research is defining those aspects that are of central importance for virtual agents. In this study, the dramaturgical model of personality developed by the social constructivists is used to delineate the domain of artificial personality. According to the social constructivists, personality is the product of three perspectives: that of the actor (the expression of psychological personality), that of the observer (the perception and interpretation of personality), and that of the self-observer (the management of self-presentations).

Most research in artificial personality has focused on the actor. This study explores the observational perspectives by considering the physical personality of the actor, defined as comprising those aspects of appearance that give rise to an initial impression of personality. It is argued in this study that modeling the impressions of physical personality would provide virtual agents not only with a means of perceiving physical personality but also with a means of creating their own socially intelligent embodiment.

To illustrate the feasibility of modeling physical personality, a study focused on modeling the trait impressions of the face using an autoassociative neural network or, equivalently, Principal Component Analysis (PCA) is presented. The performances of three-class and two-class PCAs, trained to match human classification of faces in terms of perceived dominance, masculinity, sociality, adjustment, warmth, trustworthiness, facial maturity, and gender, are analyzed. It is found that the PCAs perform better than chance, with two-class PCAs outperforming three-class PCAs. The study concludes by reporting on an investigation designed to gauge the potential of synthesizing faces with a high probability of producing specific trait impressions from within the PCA trait space.