How to Avoid "Thinking on Your Feet"
Lloyd G. Greenwald · 1995
The common otion of thinking on your feet refers to the ability to quickly make decisions in response to changing situations. Implied in this concept is that reasoning about taking action takes place at (or very near) the time of action. In this note we argue that such a skill is neither practical nor necessary in general. A key research issue concerning approaches to taking time-critical action in dynamic, uncertain environments is to determine when it is most effective to reason about taking action. Over the last decade there has been a shift away from traditional AI reasoning and representation when dealing with systems that must take time-critical actions in dynamic, uncertain environments. Many approaches are converging toward a theory of action that adopts feedback-based paradigms in order to achieve purposeful reactive behavior. One of the primary difficulties that defines the differences between alternative approaches is the tension between anticipating all possible contingencies and deliberating quickly in timecritical situations. Some have argued that what is at issue is a classic space-time tradeoff. This implies a constant level of resource expenditure that can be divided between pre-compiling and storing actions prior to execution and reasoning about actions during execution. We argue here that the real issue is more subtle than that. The issue is not only how much time to spend reasoning about actions but also when to spend that time and what information to bring to bear on the problem. Strategic use of time can lead not only to a shifting of resources between execution-time computation and storage but to an efficient optimization of overall resource requirements. The focus of our work is to take advantage of structure in the problem domain and make use of off-line meta-reasoning. The structure we require is: a model that can be used for probabilistic prediction of how the