Affective reasoning holds great potential for interactive digital entertainment, education, and training. Incorporating affective reasoning into the decision-making capabilities of interactive environments could enable them to create customized experiences that are dynamically tailored to individual users ’ ever changing levels of engagement, interest, and emotional state. Because physiological responses are directly triggered by changes in affect, biofeedback data such as heart rate and galvanic skin response can be used to infer affective changes. However, biofeedback hardware is intrusive and cumbersome in deployed applications. This paper proposes an inductive framework for automatically learning models of users’ physiological response from observations of user behaviors in interactive environments. These models can be used at runtime without biofeedback hardware to continuously predict users ’ physiological state directly from situational context in the interactive environment. Empirical studies with induced decision tree, naïve Bayes, and Bayesian Network physiological response models suggest that they may be sufficiently accurate for practical use.

Affective reasoning plays an increasingly important role in cognitive accounts of social interaction. Humans continuously assess one another's situational context, modify their own affective state accordingly, and then respond to these outcomes by expressing empathetic behaviors. Synthetic agents serving as companions should respond similarly. However, empathetic reasoning is riddled with the complexities stemming from the myriad factors bearing upon situational assessment. A key challenge posed by affective reasoning in synthetic agents is devising empirically informed models of empathy that accurately respond in social situations. This paper presents CARE, a data-driven affective architecture and methodology for learning models of empathy by observing human-human social interactions. First, in CARE training sessions, one trainer directs synthetic agents to perform a sequence of tasks while another trainer manipulates companion agents ’ affective states to produce empathetic behaviors (spoken language, gesture, and posture). CARE tracks situational data including locational, intentional, and temporal information to induce a model of empathy. At runtime, CARE uses the model of empathy to drive situationappropriate empathetic behaviors. CARE has been used in a virtual environment testbed. Two complementary studies investigating the predictive accuracy and perceived accuracy of CAREinduced models of empathy suggest that the CARE paradigm can provide the basis for effective empathetic behavior control in embodied companion agents. 1.

Abstract. Affective reasoning has been the subject of increasing attention in recent years. Because negative affective states such as frustration and anxiety can impede progress toward learning goals, intelligent tutoring systems should be able to detect when a student is anxious or frustrated. Being able to detect negative affective states early, i.e., before they lead students to abandon learning tasks, could permit intelligent tutoring systems sufficient time to adequately prepare for, plan, and enact affective tutorial support strategies. A first step toward this objective is to develop predictive models of student frustration. This paper describes an inductive approach to student frustration detection and reports on an experiment whose results suggest that frustration models can make predictions early and accurately. 1

Humans continuously assess one another’s situational context, modify their own affective state, and then respond based on these outcomes through empathetic expression. Virtual agents should be capable of similarly empathizing with users in interactive environments. A key challenge posed by empathetic reasoning in virtual agents is determining whether to respond with parallel or reactive empathy. Parallel empathy refers to mere replication of another’s affective state, whereas reactive empathy exhibits greater cognitive awareness and may lead to incongruent emotional responses (i.e., emotions different from the recipient’s and perhaps intended to alter negative affect). Because empathy is not yet sufficiently well understood, it is unclear as to which type of empathy is most effective and under what circumstances they should be applied. Devising empirically informed models of empathy from observations of “empathy in action ” may lead to virtual agents that can accurately respond in social situations. This paper proposes a unified inductive framework for modeling parallel and reactive empathy. First, in training sessions, a trainer guides a virtual agent through a series of problem-solving tasks in a learning environment and encounters empathetic characters. The proposed inductive architecture tracks situational data including actions, visited locations, intentions, and the trainer’s physiological responses to generate models of empathy. Empathy models are used to drive runtime situation-appropriate empathetic behaviors by selecting suitable parallel or reactive empathetic expressions. An empirical evaluation of the approach in an interactive learning environment suggests that the induced empathy models can accurately assess social contexts and generate appropriate empathetic responses for virtual agent control. Categories and Subject Descriptors

Abstract. Self-efficacy is an individual’s belief about her ability to perform well in a given situation. Because self-efficacious students are effective learners, endowing intelligent tutoring systems with the ability to diagnose selfefficacy could lead to improved pedagogy. Self-efficacy is influenced by (and influences) affective state. Thus, physiological data might be used to predict a students ’ level of self-efficacy. This paper investigates an inductive approach to automatically constructing models of self-efficacy that can be used at runtime to inform pedagogical decisions. In an empirical study, two families of self-efficacy models were induced: a static model, learned solely from pre-test (non-intrusively collected) data, and a dynamic model, learned from both pretest data as well as runtime physiological data collected with a biofeedback apparatus. The resulting static model is able to predict students ’ real-time levels of self-efficacy with reasonable accuracy, while the physiologically informed dynamic model is even more accurate. 1

Embodied conversational agents (ECA) have potential as facilitators for health interventions. However, their utility is limited as long as people must sit down in front of a computer to access them. This paper describes a project that is deploying an ECA on a handheld computer, and using it to assist in a psychosocial intervention aimed at providing training in problem solving skills. The agent is based upon the virtual trainer/counselor in the pedagogical drama Carmen’s Bright IDEAS, adapted for handheld use and for interaction with a human caregiver. The system will go into clinical trails in August of 2004. The paper discusses the design and technical issues involved in the transition from laptop computer to handheld device and from 3rd-person view to first-person interaction, and the plan for evaluation. The clinical trial is designed both to evaluate psychosocial outcomes and to assess user preferences in ECA interaction modalities over the course of multiple sessions of use.

Abstract Human teaching strategies are usually inferred from transcripts of face-to-face conversations or computer-mediated dialogs between learner and tutor. However, during natural interactions there are no constraints on the human tutor’s behavior and thus tutorial strategies are difficult to analyze and reproduce in a computational model. To overcome this problem, we have realized a Wizard of Oz interface, which by constraining the tutor’s interaction makes explicit his decisions about why, how, and when to assist the student in a computer-based learning environment. These decisions automatically generate natural language utterances of different types according to two “politeness” strategies. We have successfully used the interface to model tutorial strategies. 1

The use of Hofstede’s cultural dimensions

Abstract. Narrative-centered learning environments hold much promise for education and training. Much of the appeal of narrative-centered learning lies in the belief that narrative context provides a meaningful structure integrating pedagogical objectives into a unifying, coherent form that serves as a powerful

Affect has been the subject of increasing attention in cognitive accounts of learning. Many intelligent tutoring systems now seek to adapt pedagogy to student affective and motivational processes in an effort to increase the effectiveness of tutorial interaction and improve learning outcomes. To this end, recent work has begun to investigate the emotions experienced during learning in a variety of environments. In this paper we extend this line of research by investigating the affective transitions that occur throughout narrative-centered learning experiences. Further analysis differentiates the likelihood of affective transitions stemming from pedagogical agent empathetic responses to student affect.