In our current research, embodied in a large AI program called the Affective Reasoner, we simulate simple worlds populated with with agents capable of responding ``emotionally'' as a function of their concerns. Agents are given unique pseudo-personalities modeled as both a set of appraisal frames representing their individual goals, principles, preferences, and moods, and as a set of channels for the expression of emotions. Combinations of appraisal frames are used to create agents' interpretations of situations that unfold in the simulation. These interpretations, in turn, can be characterized by the simulator in terms of the eliciting conditions for emotions. As a result, in some cases agents ``have emotions,'' which then may be expressed in ways that are observable by other agents, and as new simulation events which might perturb future situations [Elliott1992]. Additionally, agents use a case-based heuristic classification system (based on [Bareiss1989]) to reason about the emotions other agents are presumed to be having, and to form representations of those other agents' personalities that will help them to predict and explain future emotion episodes involving the observed agent [Elliott & Ortony1992, Elliott1992].
Ortony, et al. [Ortony, Clore, & Collins1988] discuss twenty-two emotion types based on valenced reactions to situations being construed as goal-relevant events, acts of accountable agents, or attractive or unattractive objects (including agents interpreted as objects). This theory has been extended to include the two additional emotion types of love and hate [Elliott1992]. See figure 1.
Additionally, using the work of Ortony, et al. [Ortony, Clore, & Collins1988] as a guide, we analyzed a set of descriptions of emotion eliciting situations and created a modified set of emotion intensity variables to explain the causes of varying emotion intensity, within a coarse-grained simulation paradigm {[Elliott & Siegle1993]. We reduced the resulting set of variables to a computable formalism, and represented sample situations in the Affective Reasoner. We then isolated three areas of the simulation where variables in either the short-term state of an agent, the long-term disposition of an agent, or the emotion-eliciting situation itself, helped to determine the intensity of the agent's subsequent affective state. For each area there is an associated group of variables. The first group, simulation-event variables, comprises variables whose values change independently of situation interpretation mechanisms. The second group, stable disposition variables, consists of variables that are involved in an agent's interpretation of situations, tend to be constant, and help to determine an agent's personality and role in the simulation. We felt that, for the purposes of implementation, the distinction between these two groups was underspecified in the work of Ortony, et al. [Ortony, Clore, & Collins1988]. The last group, mood-relevant variables, contains those variables that contribute to an agent's mood state. In all there are approximately twenty such variables, although not all apply to each emotion category.
For example, the variable blameworthiness-praiseworthiness might roughly be described as the degree to which an observing agent interprets an observed agent as having upheld or violated one of the observing agent's principles, in some situation. It is derived from a set of simulation values, which might include values for the amount of effort expected in a given situation, the accountability of an agent as determined by role, and so forth. It has no default value, being determined entirely by one agent's construal of the simulated act of an another agent.
Our work has focused primarily on the detailed working out of a computational method for representing the antecedents and expression of human emotion in diverse human social situations. This has included the analysis, within the constraints of the underlying emotion theory, of many hundreds of informally described social situations which have given rise to emotions. To date the computational mechanism has included, among other components, (a) the construction of hundreds of appraisal frames, which include slots for reasoning about intenstiy and mood, in domains as diverse as, for example, financial accounting, stories, playing poker, and sales, (b) pseudo personality types made up of these appraisal frames, (c) the use of these pseudo personalities for construing situations with respect to the concerns of simulated agents, thus giving rise to ``emotion generation'' in simulation runs, (d) the generation of simple emotion instances based on the twenty-four emotion categories, (e) the generation of actions through approximately 450 channels (about twenty for each emotion category) consistent with the simulated emotions they are intended to express (each of which may, in turn, contain multiple manifestation instances), (f) abductive reasoning about the emotions expressed by other agents in the system, (g) the internal representation of the presumed pseudo personalities of observed agents by observing agents, (h) simple ``explanations'' of the emotions with respect to their emotional antecedents within the simulation, (i) simple models of relationship between the agents, allowing for ``emotions'' based on the concerns of others, and (j) the inclusion of the user as one of the agents in the system about which reasoning may take place.
Most recently we have been working on opening up communication channels with the user of the system through the addition of modules for speech recognition, inflected speech generation, indexed music-on-demand (using a midi interface and a 400 voice Proteus MS-PLUS synthesizer), and facial expression for the agents.
The broad long-range goals we would like to see pursued include number of applications we envision as made possible by the representational capabilities of a system such as this. Among these are the building of a computer that has some capability of categorizing, and responding to, a user's affective state; the building of systems that allow users to interactively explore the emotional content of a broad range of simulated social situations (e.g., for tutoring, for socially unskilled psychotherapy patients, and for military stress applications); testing the use of ``emotionally aware'' automated software agents as a way of enhancing the user's engagement in educational, and other software; using emotionally aware agents to communicate priorities naturally to the user, such as with an automated assistant for meetings, or the communication of technical concerns to less technical users through the focus of attention (e.g., operating systems, financial analysis), the use of computer-based emotion expression as an authoring tool (e.g., as online feedback for students), the construction of games that include a complex, yet cohesive, emotion and personality component; the use of the underlying emotion theory to analyze, and manipulate, the automated telling of stories; and platforms for testing the link between music and emotion expression.
Figure: Emotion types (Table based on [O'Rorke and Ortony, 1992] and [Elliott, 1992])