Role of Emotions in Rational Agent Design: General Principles and Applications

This research is predicated on the thesis that the notions of emotions and feelings can be formalized and be made useful in designing artificial intelligent agents that are to interact with a complex uncertain environments, possibly populated by other agents and humans. Our formalization starts from the formal description of rational agent design based on decision theory, according to which agents act so as to maximize the expectation of their performance measure. Interestingly, the decision-theoretic model of decision making can be used to formally define the emotional states and feelings of a rational agent. Our definitions show how emotional states transform the agent's decision-making situation, say, by making the agent more short-sighted, by altering the agent's subjective performance measure, or by modifying the probabilities that the agent assigns to states of the world.

Having the formal definitions of emotional states allows us to show how, and why, emotions may be useful to a rational agent. First, by modifying the decision-theoretic model used for deliberative rationality, emotions allow the agent to control the allocation of its cognitive resources and the complexity of its deliberations under time pressure in an uncertain environment. For example, limiting the number of alternative behaviors considered, or shortening the time horizon of these alternatives allows for more rapid decision-making.

Second, emotions and feelings turn out to be valuable when the agent finds it useful to inform other agent(s) about its own internal state. Instead of having to describe the details of its internal state, and running the risk of being misunderstood if the other agents are engineered differently, the agent can use more abstract and universal terms. For example, notions of stress or panic may be convenient to express the fact that the urgency of the situation forced the agent to look at only short-term effects of its actions. Thus, in the context of communication with other agents, the formal definitions of emotions serve as semantics of emotional terms, with which the agents can express their own internal states and understand the states the other agents are in.

Third, well-defined emotional states of self and others are crucial in the agent's interaction with humans. Frequently, human-computer interaction is impeded by the machine being hopelessly out-of-step with the emotional state of the human user. It turns out, however, that the users' emotional state can be assessed using measurable and inferred factors, and it is important for the machine to model the effect the user's emotional state has on the user's decision-making and his/her tendency to action.