Becoming Increasingly Reactive

augmenting its reactive component whenever it is forced to plan. When used to control a laboratory mobile robot, the We describe a robot control architecture which Theo-Agent in simple cases learns to reduce its reaction combines a stimulus-response subsystem for rapid time for new tasks from several minutes to less than a reaction, with a search-based planner for handling second. unanticipated situations. The robot agent continually The research reported here is part of our larger effort chooses which action it is to perform, using the stimulustoward developing a general-purpose learning robot response subsystem when possible, and falling back on the architecture, and builds on earlier work described in planning subsystem when necessary. Whenever it is [Blythe and Mitchell 89]. We believe that in order to forced to plan, it applies an explanation-based learning become increasingly successful, a learning robot will have mechanism to formulate a new stimulus-response rule to to incorporate several types of learning: cover this new situation and others similar to it. With • It must become increasingly correct at predicting experience, the agent becomes increasingly reactive as its the effects of its actions in the world. learning component acquires new stimulus-response rules that eliminate the need for planning in similar subsequent • It must become increasingly reactive, by reducing situations. This Theo-Agent architecture is described, and the time required for it to make rational choices; results are presented demonstrating its ability to reduce that is, the time required to choose actions routine reaction time for a simple mobile robot from consistent with the above predictions and its goals. minutes to under a second. • It must become increasingly perceptive at distinguishing those features of its world that impact its success.