Juan Velazquez at the MIT Artificial Intelligence lab has developed a computational model called Cathexis for generating emotions in autonomous agents. It makes use of a set of about six emotion families, such as anger, and fear, drawing from the work of Ekman, Izard, and others, as well nine motivational drives, such as hunger, fatigue, and curiosity, based on control systems theory. Each of these has releasors which regulate the duration, and intensity, of states over time. Additionally, the emotion states have elicitors that fall in the categories neural, sensorimotor, motivational, and cognitive.
Of particular interest in this system is that it has a fairly robust model of the ebb and flow of affective states over time. To understand how difficult a problem this is, consider the aforementioned state of fear. Suppose that one is fearful about a menacing person believed to be immediately threatening one's social status. Certainly it is true that the more important the goals, or the more likely that they will be blocked, the more intense the fear. But now consider one's dispositional fear of death. By contrast with a fear of losing social status, this latter fear is likely to be more intense, on the one hand, because of the extreme nature of death, but, on the other hand, less intense because it is less immediate. Now we have to ask, is this long-term, elemental, fear (a) more intense on average (i.e., a stronger influence on current behavior), (b) more intense for short, repeating, bursts, (c) less intense at each moment (and of course, what is a moment in an AI system?) but more intense in sum, (d) etc.? Is intensity strictly quantitative, or is it qualitative? (That is, is being mildly pleased closer in nature to being slightly angry, or to being intensely rapturous?). The Velasquez model goes farther in addressing these difficult issues than many through its computational mechanisms of elictors and releasors.
The integrated behavior systems, like some of those discussed by researchers in the Sloman group, operate somewhat autonomously, competing in an inhibition network for chances to manifest themselves. Other behaviors, as well as initial effectors, can regulate a behavior's form and eligibility to fire. In this model a winner-take-all strategy, which is somewhat controversial in its pure form, is used.
The model is comprehensive and designed for use in both software agents, and robots. It has been partially implemented in synthetic agents such as Simon the Toddler, and a robot Mutant [Fujita & Kageyama1997]. Only time will tell if the sophisticated checks and balances used to create the dynamic processes in the system prove generally applicable in practice, but the group is to be commended on the push toward implementation and testing in such varied environments.