Topic and Approach

The purpose of the present work is to characterize the behavior pattern of deliberated retry in logicist terms. Deliberated retry is where an agent pursues goals by trying actions or action sequences that are likely to achieve the goal, and where the agent responds to failure by trying an alternative plan or action sequence. (We identify plans with action sequences). The paper proposes a rst-order theory using a narrative time-line approach in which each model is a history of the world where the agent's successive actions exhibit deliberated retry. This work is intended to be used in the design of autonomous agents, and in particular in our WITAS project which is concerned with the control system of an intelligent airborne vehicle (UAV). A logic of this kind must of course build on existing work in logics of actions and change, but it also imposes some speciic requirements on the host logic. Besides the characterization of deliberated retry, which is addressed in this article, it must also characterize the success and failure of actions, it must account for exogenous events that innuence the execution of an action, and it must allow actions to be described on diierent levels of resolution. The present section discusses the integration of these diierent aspects of the entire agent architecture on the logic level. 1.1 Goals vs high-level actions We shall consider goals as the end states of particular kinds of actions, namely goal-directed actions. Therefore, the question of characterizing goal-directed behavior is reduced to the question of how goal-directed actions are decomposed into procedural ones. By this approach we avoid the need to introduce goals as a separate type of entities, and we introduce the possibility (although it will not be used in the present paper) of having goals and actions on several levels. A goal-directed action is then viewed as an open-ended process: the robot sets out to achieve a certain goal, for example to nd and retrieve a particular object, or to obtain more fuel. It chooses certain lower-level actions which are likely to achieve the goal, but these lower actions may succeed or fail; if they fail then the robot diagnoses the error and tries to achieve the same goal in some other way. At the lowest level in this structure one nds the procedural actions, but above them there is a structure of goal-directed actions, possibly on several levels. They are …