1 Plans and Behavior in Intelligent Agents

An "intelligent agent" is a versatile and adaptive computer system. It performs diverse cognitive and physical behaviors in its efforts to achieve multiple goals in a dynamic, uncertain task environment. In this paper, we address the question: How should plans influence the cognitive and physical behavior of intelligent agents? In contrast to the well-known model of plans as executable programs, we propose that intelligent agents can make better use of plans that describe their intended behavior. This kind of model has been discussed by other researchers, but it has not been operationalized. We instantiate this model with operational definitions of plans, planning, and plan following. We present an agent architecture for using such plans to guide an agent's cognitive and physical behaviors. Experimental results from an office robot demonstrate important capabilities engendered by our approach: (a) coordination of diverse cognitive and physical behaviors to achieve multiple goals; (b) robust improvisation of goal-directed behavior over a range of situations; and (c) ready exploitation of asynchronously acquired knowledge and behavioral capabilities. An "intelligent agent" is a versatile and adaptive computer system that performs diverse cognitive and physical behaviors in its efforts to achieve multiple goals in a dynamic, uncertain task environment. A typical agent performs several interacting tasks. For example, an agent's cognitive tasks include assessing its present situation to decide what specific goals to pursue, planning an appropriate goal-directed course of action, explaining its goals and plans to other agents, etc. Physical tasks include perceiving features of its world needed to perform a cognitive or physical task, perceiving important unexpected features of its world, moving about in its world, etc. Agents often have alternative methods for performing a given task. For example, an agent might plan a goal-directted course of action by deductive or case-based methods. It might move about in its world by motion planning or pure feedback-control methods. In a given situation, the agent must decide which tasks to perform and by which methods to perform them. It must make these decisions repeatedly in the context of a dynamic task environment. In the course of its experience, an agent acquires knowledge and capabilities that allow it to perform new tasks or perform known tasks in new ways. In this paper, we address the question: How should plans influence the cognitive and physical behavior of intelligent agents? 2. What is a plan? A plan can take different forms, depending on …