Dynamic Knowledge Integration during Plan Execution

The goal of our work is to develop architectures for general intelligent agents that can use large bodies of knowledge to achieve a variety of goals in realistic environments. Our e orts to date have been realized in the Soar architecture. In this paper we provide an overview of plan execution in Soar. Soar is distinguished by its use of learning to compile planning activity automatically into rules, which in turn control the selection of operators during interactions with the world. Soar's operators can be simple, primitive actions, or they can be hierarchically decomposed into complex activities. Thus, Soar provides for fast, but exible plan execution. Following our presentation of plan execution, we step back and explicitly consider the properties of environments and agents that were most in uential in Soar's development. From these properties, we derive a set of required general agent capabilities, such as the ability to encode large bodies of knowledge, use planning, correct knowledge, etc. For each of these capabilities we analyze how the architectural features of Soar support it. This analysis can form the basis for comparing di erent architectures, although in this paper we restrict ourselves to an analysis of Soar (but see Wray et al. (1995) for one such comparison). Of the capabilities related to plan execution, one stands out as being at the nexus of both the environment/agent properties and architecture design. This is the capability to integrate knowledge dynamically during performance of a task. We assume that central to any agent is the need to select and execute its next action. To be intelligent, general, and exible, an agent must use large bodies of knowledge from diverse sources to make decisions and carry out actions. The major sources of knowledge include an agent's current sensing of the world, preprogrammed knowledge and goals, the agent's prior experience, instructions from other agents, and the results of recent planning activities (plans). However, many plan execution systems base their decisions solely on their plans, thus limiting their exibility. One reason is that it is di cult to integrate plan control knowledge dynamically and incrementally with all of the other existing knowledge as the agent is behaving in the world. What we hope to show is that this is an important capability for responding to realistic environments, and that Soar's architectural components provide this capability in a general and exible way. Soar has been used for a variety of real and simulated domains, including real and simulated mobile robot control [Laird and Rosenbloom, 1990], real and simulated robotic arm control [Laird et al., 1991], simulated stick-level aircraft control [Pearson et al., 1993], simulated tactical aircraft combat [Tambe et al., 1995], and a variety of other simulated domains [Covrigaru, 1992, Yager, 1992]. We will use a variety of examples from these domains to illustrate our points through-