CENTER FOR INTERDISCIPLINARY RESEARCH 2004 FACULTY FELLOWSHIP APPLICATION FORM Name Campus Address Campus Phone Email School Department Title of Proposal Research Grant Amount Requested Identify which Disciplines Involved Stephen J. Read, Ph.D. 821 SGM, MC – 1061 740-2291 read@usc.edu College of Letters, Arts, and Sciences Department of Psychology Computational Models of Personality for Intelligent Agents in Learning and Tutoring Systems $42,493.00 Psychology, Communications, Neuroscience, Computer Science List all key faculty involved: 1. 2. 3. Name Lynn C. Miller, Ph.D. Kwan Min Lee, Ph.D. Larry Swanson, Ph.D. 4. Stacy Marsella, Ph.D. 5. Hannes Vilhjalmsson, Ph.D. 6. Wayne Zachary, Ph.D. Department Professor, Annenberg School for Communication Assistant Professor, Annenberg School for Communication Milo Don and Lucille Appleman Professor of Biological Sciences, Head of the NIBS program Research Scientist, USC Information Sciences Institute, The Center for Advanced Research in Technology for Education Research Scientist, USC Information Sciences Institute, The Center for Advanced Research in Technology for Education President, CHI systems, Philadelphia, PA. Developers of cognitive modeling software. ________________________________ Applicant’s Signature Gerald C. Davison, Chair of Psychology Print Chair’s Name Donal Manahan, Dean of Research, College of LAS Print Dean’s Name ________________________________ Date ________________________________ Chair’s Signature ________________________________ Dean’s Signature Stephen J. Read, Ph.D. Computational models of Personality Aims The intent of this proposal is to further develop a computational model of human personality. Our purposes are twofold. The first is to integrate information from a number of different domains to develop a greatly improved theory of human personality, in the form of an explicit computational model. The second purpose is to use this computational model to create computer based, intelligent agents with recognizable personalities that can be used as tutors and virtual peers in highly engaging, socially sensitive virtual learning systems for children and adults. The model will integrate work from a variety of disciplines: work from personality psychology on the structure of human personality and temperament, work from neuroscience on the neurobiology of temperament and motivation, work from psychology, communication, and computer science on the behavioral expression of personality, and work from computer science and cognitive science on Intelligent Agents. The two purposes of this project will mutually support and inform each other. Not only will integrating the work from psychology, communication and neuroscience help form the basis of a computational model of personality for Intelligent Agents, but at the same time building the model and using it as the basis for the personality of an Intelligent Agent should lead to a number of interesting questions and greatly inform our knowledge of the nature of human personality. Further, developing intelligent agents will provide a strong set of constraints on our model and provide a strong test of its plausibility. Importance of this work This work has both theoretical and practical importance. From a theoretical point of view, we are developing a computational model of human personality that will provide a detailed process model of the dynamics of the underlying motivational and neurological structures. Because such a detailed, integrated process model does not currently exist, this model should make a strong theoretical contribution to the understanding of human personality. From a practical point of view, the model would make a major contribution to creating more socially engaging and effective Intelligent Agents. Such agents are becoming increasingly widespread. They may serve as a tutor in an educational system or as a collaborator in a virtual work group. Unfortunately, one uniform failing of current agents is that they either fail to display realistic human personality or they display a particular, hand tuned personality. There is no way to create the underlying structures needed for an Intelligent Agent with an easily customizable personality. Yet, there are numerous reasons why being able to do so is important. For example, science educators have become increasingly interested in developing virtual learning environments in which a child’s learning is aided by an animated guide or tutor (see Forbus and Feltovich (2001) for a recent review) or in which the child is part of a work group of virtual peers. In an attempt to push this work forward, we are 1 Stephen J. Read, Ph.D. Computational models of Personality developing a Science of Learning Center whose central focus is creating virtual learning systems that are sensitive to the emotional and motivational state of the learner and that can adjust their behavior so as to maximize the child’s engagement and motivation to learn. In such systems, Intelligent Agents would serve as tutors, as well as being part of virtual work groups in which the learner would participate in a group activity. Engaging characters should increase the child’s involvement and guide and scaffold the child’s performance. To do so these agents need interesting and engaging personalities. Further, there is evidence that the specific personality traits of the tutor should impact the children’s response to the tutor and the tutor’s effectiveness. However, the ability to create engaging characters, and particularly the ability to tailor their personalities to maximize student interest and learning, is hobbled by the lack of a computational model of personality that allows one to quickly and easily modify agents’ personalities. Intelligent Agents are also playing an increasing role in training systems for the military and for others. Personality is important in various ways in these systems. First, there is an increasing concern with the necessity of training soldiers to deal with a variety of different personalities and to deal with people from an array of different cultures. Second, in group training simulations, trainees need to learn to deal with group members’ personalities (e.g., dominant vs. agreeable). Third, agents with realistic personalities should greatly increase user involvement with the training simulation. Fourth, Intelligent Agents with realistic personality structures may be easier to develop, easier to control and interact with, and more robust than those built with existing technology. A personality-based architecture could provide the developer with familiar, general dimensions for control (for example, adjustment of a ‘helpfulness’ trait may dramatically simplify the problem of obtaining the right balance between 'picking up' and 'dropping' tasks by an Intelligent Agent). In addition, the personality framework will provide users with a more interpretable basis for interacting with Intelligent Agents. Theoretical Approach We will take two parallel, mutually supporting approaches to the development of our computational model. First, we will integrate findings on the components and processes underlying personality and temperament with existing symbolically based, computational architectures for cognition to create a Personality - enabled Architecture for Cognition (PAC). The research will integrate symbolically- and biologically-based theories of personality to produce a hybrid cognitive/neural/biological framework that provides a generative model of personality for an intelligent agent. Some of this work will be done with Stacy Marsella and Hannes Vilhjalmsson at USC’s Information Sciences Institute (ISI). Another part of this work will be done with CHI Systems and will directly support the design, implementation, and evaluation of a new cognitive architecture for Intelligent Agents. 2 Stephen J. Read, Ph.D. Computational models of Personality A suitable personality infrastructure will provide a rich skeleton of motivation on which the task specific behaviors of Intelligent Agents can be developed. Currently, lacking such a personality/ motivational foundation, Intelligent Agent developers face difficult challenges in adjusting the relative priorities of the many goals that are required to address diverse required agent functions. As Intelligent Agents’ roles become more complex and more robustness is required, the effective coordination of model components is especially difficult to achieve. At the same time, we will continue work on our previous neural network model of personality (Read & Miller, 2002) in collaboration with Larry Swanson in the NIBS program. As a connectionist model it has some advantages for theory development: it can be more easily tied to brain structures and their principles of operation, and it is easier to experiment with different assumptions and structures. Personality Framework. Our personality framework will integrate results from multiple sources, ranging from psychometric work on trait scales and lexical analyses of trait language to recent work in neuroscience, identifying specific brain systems for motivational domains. Our goal is an articulated, general model in which personality is based on hierarchically organized motivational systems, ranging from individual goals to higher order approach and avoidance systems. We represent personality traits as individual differences in goal-based structures. These differences affect an agent’s interpretation of a situation, and the agent’s responses to the behavior of other agents. Recent work by the research team and others suggests that we can effectively capture personality by differential configurations of goals, plans, resources, and beliefs (e.g., Miller & Read, 1991; Read, Jones, & Miller, 1990; Read & Miller, 2002). Recent work in neuroscience and temperament suggests that human motivational systems are organized into at least two levels. Mapping of brain circuits and evolutionary analyses provide evidence for a set of level one motivational systems that handle the variety of major adaptive problems that people must deal with in everyday life, such as social bonding, development of authority relations in groups, exploration and play, mating and parenting, and self-preservation and physical safety. Each of these motivational systems organizes sets of individual goals, which are the basis of specific traits. For example the social bonding system might organize the goals related to such traits as ‘helpful’, ‘friendly’, and ‘dutiful’. At a higher level are two overarching motivational systems -- a behavioral approach system (BAS) which governs sensitivity to reward (central to Extroversion) and approach to rewarding stimuli, and a Behavioral Inhibition System (BIS), which governs sensitivity to punishment (central to Neuroticism) and avoidance of threatening stimuli. Recent neuropsychological research suggests that these systems are mapped into the left and right Prefrontal Cortex, respectively, and may integrate and provide a “read-out” from the lower level motivational systems. These two levels are highly and bi-directionally interconnected. Thus, global changes in the BAS and BIS can 3 Stephen J. Read, Ph.D. Computational models of Personality influence the behavior of all the lower level goal systems that are integrated by the higher level system. Changes in the lower level motivational systems can influence all of the individual goals that are organized by that system, as well as providing inputs to the BAS and BIS. Expression of personality. Large literatures in psychology, communications, sociology, and anthropology have shown how personality is communicated; through body orientation, gaze direction, tone and speed of speech, the form of speech acts, and choices of behaviors. This literature will provide a foundation for behaviors of our agents. I will collaborate with Stacy Marsella, Hannes Vilhjalmsson, and Kwan Lee in implementing these behaviors in Intelligent Agents. Stacy Marsella is using this literature to help build emotionally expressive Intelligent Agents and Hannes Vilhjalmsson’s research focuses on the role of nonverbal cues in face-to-face interaction and how these cues can be autonomously generated in interactive animated characters based on linguistic and social context. Further, Kwan Lee, has begun to use this literature in examining people’s responses to “personality cues” in Intelligent Agents. For instance, Nass and Lee(2001) showed that the perception of extroversion in a computer-synthesized voice is influenced by number of words per minute, loudness of the voice, fundamental frequency, and degree of frequency variation. Program of Work and Proposed Products. During the year of the fellowship I would do the following. First, I would work with Larry Swanson and others on increasing my knowledge of the neuroscience of motivational systems. Second, I would work with Lynn Miller and Larry Swanson on further developing our neural network model of personality. Third, I would collaborate with Stacy Marsella, Hannes Vilhjalmsson, Lynn Miller, Kwan Lee, and CHI Systems to develop a computational architecture for personality in Intelligent Agents and to work on building socially engaging agents that exhibit realistic human personality. This work will greatly strengthen the foundations for the Socially Engaging Agents component of our resubmission of our NSF Sciences of Learning Center grant. Finally, I intend to submit research grant proposals to two different groups. One set of proposals will be sent to Federal agencies that support the development of Intelligent Agents, such as the NSF program on Information Technology Research (ITR) and military agencies, such as ONR, DARPA, and AFOSR, that are interested in Intelligent Agents in training systems. (I currently am part of a funded proposal from AFOSR with CHI systems to develop a personality enabled cognitive architecture.) A second proposal will be sent to the NIMH or NSF programs on neuroscience to support work on the computational model. References Cassell, J., & Bickmore, T. (in press). Negotiated Collusion: Modeling social language and its relationship effects in intelligent agents. User Modeling and Adaptive Interfaces. 4 Stephen J. Read, Ph.D. Computational models of Personality Forbus, K. D., & Feltovich, P. J. (2001). Smart Machines in Education: the coming revolution in educational technology. Cambridge, Mass: MIT Press. Miller, L. C., & Read, S. J. (1991).On the coherence of mental models of persons and relationships: A knowledge structure approach. In G. J. O. Fletcher & F. Fincham (Eds.), Cognition in Close Relationships. (pp. 69-99). Hillsdale, NJ: Erlbaum. Nass, C., & Lee, K. M. (2001). Does computer-synthesized speech manifest personality? Experimental tests of recognition, similarity-attraction and consistency-attraction. Journal of Experimental Psychology: Applied, 7, 171181. Read, S. J., Jones, D. K., & Miller, L. C. (1990). Traits as goal-based categories: The importance of goals in the coherence of dispositional categories. Journal of Personality and Social Psychology, 58, 1048-1061. Read, S.J., & Miller, L.C. (2002). Virtual personalities: A neural network model of personality. Personality and Social Psychology Review. Background of Collaborators Lynn C. Miller and I have collaborated extensively over the last 15 years on goal-based models of personality. She is trained as a personality psychologist. Her current work focuses on computational models of personality, using Interactive Video to change risky sexual behavior, and on evolutionary and biological bases of human mating. Larry Swanson is a world expert on the structure of mammalian motivational systems. Stacy Marsella is a Research Scientist at USC’s ISI. He is an expert in emotional and cognitive modeling for Intelligent Agents and is intensively involved in the construction of emotionally expressive Intelligent Agents. Hannes Vilhjalmsson is a Research Scientist at USC’s ISI. His research focuses on the role of nonverbal cues in face-to-face interaction and how these cues can be autonomously generated in interactive animated characters based on linguistic and social context. Wayne Zachary is the President of CHI Systems, which manufactures one of the premier software systems for the construction of computational models of cognition and has extensive experience in building computational models of cognition and training systems that rely on intelligent agents. We have recently received funding from AFOSR to explore the integration of personality into their architecture for programming cognitive systems. Kwan Min Lee’s work has focused on how personality is expressed through various characteristics of an Intelligent Agent, such as through vocal characteristics and nonverbal behavior. He received his Ph.D. from Cliff Nass at Stanford and continues this work at USC. 5 Stephen J. Read, Ph.D. Computational models of Personality Budget Item Amount Psychology Research Assistant Stipend 50% for 12 months $21,333 Psychology Research Assistant Tuition (16 units) $16,000 Xeroxing, materials and supplies $2,000 PowerMacintosh G5 computer for modeling $3,160 Total $42,493.00 Budget Justification Development of the computational model will depend upon extensive library research on what is currently known about the structure of personality, the bases of temperament, the neuroscience of motivational systems, and how personality is behaviorally expressed. It will also require becoming and staying up to date on what is happening in the burgeoning field of work on what are variously called Intelligent Agents, Autonomous Agents, or Cognitive Agents. Much of this library research and the synthesis of this research will be carried out by an advanced psychology graduate student. In addition, a central part of this project will be the further development of our neural network model of personality. Programming of this model will also require the continued involvement of a graduate research assistant. 6