A Classification of Cognitive Agents

A Classification of Cognitive Agents Mehdi Dastani (mehdi@cs.uu.nl) Institute of Information and Computer Sciences P.O.Box 80.089 3508 TB Utrecht, The Netherlands Leendert van der Torre (torre@cs.vu.nl) Department of Artificial Intelligence, Vrije Universiteit Amsterdam De Boelelaan 1081a 1081 HV Amsterdam, The Netherlands Abstract In this paper we discuss a generic component of a cognitive agent architecture that merges beliefs, obligations, intentions and desires into goals. The output of belief, obligation, intention and desire components may conflict and the way the conflicts are resolved determines the type of the agent. For component based cognitive agents, we introduce an alternative classification of agent types based on the order of output generation among components. This ordering determines the type of agents. Given four components, there are 24 distinct total orders and 144 distinct partial orders of output genera- tion. These orders of output generation provide the space of possible types for the suggested component based cognitive agents. Some of these agent types correspond to well-known agent types such as real- istic, social, and selfish, but most of them are new characterizing specific types of cognitive agents. Introduction Imagine an agent who is obliged to settle his debt, desires to go on holiday, and intends to attend a conference. Suppose that he believes he can only af- ford to finance one of these activities and decides to pay his checks to settle his debt. Unfortunately, our agent does not earn much money and is in the habit of buying expensive books. Therefore, he runs again into debt after a short while. Despite the fact that he still has the same obligation, desire, and inten- tion and believes that he can only aAEord to finance one of these activities, he decides this time to attend the conference. Directly after this decision, he hears that the conference is cancelled and he receives a telephone call from his mother telling him that she is willing to pay his checks for this time. The agent is now happy and decides to go on holiday. Our agent has a friend who has the same obligation, desire, and intention, and likewise believes that he can only af- ford to finance one of these activities. In contrast to our agent, this friend decides to go on holiday. How- ever, he is late with arranging his holiday; all travel agencies are sold out. Therefore, he decides to at- tend the conference. In a diAEerent situation where these two agents are obliged to visit their mothers, desire to go to cinema, and believe they cannot do both simultaneously, the first agent decides to visit his mother while his friend goes to cinema. Yet in another situation where these agents intend to clean up their houses, are obliged to help their friends, and believe they cannot do both, they decide to clean up their houses. Although these agents behave diAEer- ently, each of them seems to follow a certain be- havior pattern under diAEerent situations. The first agent seems to be more sensitive to his intentions and obligations than to his desires while the second agent seems to prefer his desires more than his in- tentions and obligations. Moreover, the first agent seems to be indiAEerent towards his intentions and obligations while the second agent seems to prefer his intentions above his obligations. These charac- teristics and principles that govern agent’s actions and behavior determine the type of cognitive agents and can be used as the basis for a classification of cognitive agents. We are motivated by the studies of cognitive agents where the behavior of an agent is defined in terms of rational balance between its mental atti- tudes [1, 9, 5]. A classification of cognitive agent types specifies possible ways to define the rational balance. Beside the scientific need to study possible definitions of rational balance in a systematic way, a classification of cognitive agent types is important for many applications where it is impossible to spec- ify agent behavior in specific and usually unknown situations. In such applications, it is important to specify the behavior of agents in strategic terms and by means of types of behavior. In [2] we investigate the design and implemen- tation issues of generic component-based cognitive agents. In the present paper, we propose an alter- native classification of cognitive agent types. There has been many formal and informal studies propos- ing agent types [1, 8, 4]. In these studies, there is a trade-oAE between the space of possible agent types and their precise and formal definitions. In partic- ular, informal studies provide a rich space of possi- ble types of cognitive agents and ignore their precise definitions, while formal studies provide precise def- inition of agent types but ignore the richness of the space of possible types. The proposed classification of cognitive agent types in this paper is formal and in terms of a generic component based architecture.