A workable definition of computerized agents

摘要

Recently, several suggestions for (multi-) agent theories have been presented. In order to develop a meaningful formal theory it is necessary also to specify a model, i.e., how to interpret the theory. This, in turn, requires a well-defined domain, otherwise it is not possible to define predicates, etc. We argue that current suggested agent theories do not have a well-defined domain and consequently there is neither a well-defined interpretation of the used symbols. In fact, the common interpretation is more of a metaphysical character. The reason behind the failure, we believe, is that the attempts are based on a diffuse idea of the internal structure of computerized agents, often described in terms of human-like "mental states" such as beliefs, desires, and intentions. This is in fact a circle reasoning since a computerized agent is not defined and consequently cannot have an internal structure. Because of the "internal structure" idea, a sound definition of agents in these terms is necessary to get a formal theory. We show that none of the most known agent definitions fulfills this requirement, and that it is very hard, perhaps impossible, to come up with such a definition. Another problem with current so called theories is that they aim to be very general, covering every aspect of agenthood. This is too optimistic since there is obvious evidence that some aspects of the human agency are not describable. In short, the task of developing a general agent theory is simply not feasible. Instead we argue that agent theories should be relativized to particular human-like situations without reference to obscure metaphysical interpretations as internal structures of agents that are yet to be defined. We suggest that the proper approach to constructing agent theories is to take suitable social theories and transform them to technical domains where the new artificial actors become (machine) agents. In fact, there already exist successful transformations of this kind, e.g., theories of market behavior into market-oriented programming.