A Layered Environment for Reasoning about Action

An intelligent system reasons about--controls, explains, Iearns about--its actions, thereby improving its efforts to achieve goals and function in its environment In order to perform effectively, a system must have knowledge of the actions it can perform, the events and states that can occur, and the relationships among instances of those actions, events, and states. We represent such knowledge in a hierarchy of knowledge abstractions and impose uniform standards of knowledge content and representation on modules within each hierarchical level. We refer to the evolving set of such modules as the BB* environment. To illustrate, we describe selected elements of BB*: (a) the foundational BBl architecture: (b) the ACCORD framework for solving arrangement problems by means of an assembly method: (c) two applications of BBl-ACCORD, the PROTEAN system for modeling protein structures and the SIGHTPLAN system for designing construction-site layouts: and (d) two hypothetical multifaceted systems that integrate ACCORD, PROTEAN, and SIGHTPLAN with other possible BB* frameworks and applicatjons. . 1. Overview: Four Themes wHuman intelligence depends essentially on the fact that we can represent in language facts &bout our situation, our goals, and the effects of the various actions we can perform.” John McCarthy [35] “In the knowledge is the power? Edward A. Feigenbaum [14] “The fact, then, that many complex systems have a nearly decomposable, hierarchic structure is a major facilitating factor enabling us to understand, to describe, and even to ‘see’ such systems and their parts.” Herbert A. Simon [48] ’ We be&in with a premise: An intelligent system reasons about its actions. Of course, we do not mean to suggest that a system should engage in extended contemplation of every one of its computational and physical actions, but rather= (a) that it can reason about many of its actions: ’ (a) that it does reason about them much of the time; and (c) that its reasoning improves its “efforts to achieve goals and otherwise function in its environment. A system might reason about its actions in various ways and with various consequences (see Figure la) . For example, a system might control its actions: decide which actions to perform at particular points in time. Control reasoning can affect the resources the system consumes in pursuing a goat. the side effects it produces, and the probability of achieving its goal [8, 9, 13, 17, 23, 26, 271. As a second example, a system might explain’ its actions: describe the ways in which the actions it intends to perform or has performed serve its goals. Explanation typically serves social functions, such as teaching another individual how to ’ perform a task or persuading another individual that one is performing the tkk competently 15, 6, 21, 22). As a third example, a system might learn about its actions: modify its Tbility or inclination to perform particular actions in appropriate circumstances. Learning enairles the system to expand and improve its capabilities [24, 32, 33, 35, 38, 39, 451. While a system could perform many other important types of reasoning about its actions, we focus on control, explanation, and learning. L Control Intelligence) Explanation Learning (4 Figure L Four Themes. (a) An intrlllgrnt system reasons about Its actions. The BBl architecture provides knowledge structures and a basic mechanism for control. explanation, and laming (b) To pufom effectively, a system must hare knowledge abour irr actions. Frameworks explicitly represent knowledge about task-specific actions, eventx, and stata and the relationships among them. (c) Knowledge is reprcsenred In an absrractlon hierarchy. The BB. environment comprises an evolving body of knowledge: the BBl architecture, task-specific frameworks, such as ACCORD, and domain-specific applications, such as PROTEAN (see Table 1). Conversely. an application system layen application-specific knowledge on a framework, which iayen task-specific knowledge on the BBl architecture. (d) Knowledge modules within a lcrcl saris/y uniform rrandards of knowledge conrent and representation. As a consequence. BB” achieva open systems integration: Independently constructed modules can be fully integrated in implementation and reasoning.