Intelligent Architectures for Knowledge Sharing: A Soar Example and General Issues

In this talk we present a model of knowledge assessment based on architecture-specific metrics and present a method of knowledge sharing called Direct Knowledge eXchange (DKX). Although embryonic from a knowledge management perspective, the notion of DKX between homogeneous knowledge engines and knowledge metric assessment represents a small step to examining how distribution of knowledge can lead to remote “knowledge invocation on demand” types of distributed problem solving systems and practical mechanisms to assess their resource value. Summary Soar is a production system that characterizes all symbolic goal-oriented behavior as search in problem spaces and serves as an architecture for general intelligent behavior (Laird, Rosenbloom & Newell 1987). A problem space defines a set of states that can be reached within that problem space, and an associated collection of operators. Operator applications move Soar from state to state and, consequently, define search in the problem space. Decisions are the primitive acts of the system used for search (i.e., generation and selection) of appropriate problem spaces, states, and operators, as well as the application of operators for new state configuration, in the pursuit of goals specified in a goal hierarchy. To achieve problem-solving goals, Soar operates in terms of a two-phase decision cycle. Each cycle starts with an elaboration phase followed by a decision phase. Together, these phased mechanisms, coupled with an embedded set of primitive preferences, allow for problem solving to ensue through the specification of appropriate sequences of operators. Soar is impasse-driven. In situations where the operator selection cannot unambiguously proceed (e.g., via incomplete or inconsistent preferences), then an impasse occurs, and a new subgoal is established with an associated problem space, and the process recurses. Copyright © 2002, American Association for Aritificial Intelligence (www.aaai.org). All rights reserved. The resolution of a subgoal in Soar is achieved by finding knowledge that resolves higher-level impasses, allowing problem solving to proceed. When this occurs, Soar learns. Specifically, “chunks” are produced that are productions that map working memory elements defining impasse situations (as antecedent conditions) into the results of subgoals (as consequent conditions). Chunking can be viewed as a form of explanation-based learning, but it is at a level articulated in specific and uniformly applied cognitive mechanisms. Subsequent encounters with similar impasse conditions can thus be resolved more directly (and with less deliberation) with the newly acquired chunks. Soar has learned. Key to learning in general and generalized learning in particular is the ability to transfer knowledge. Soar has three basic forms of transfer: within-trial (chunks can be used as soon as they are built), between-trial (chunks are improved with repeated trials on a task), and across-task (chunks can apply to similar problems). We are exploring another form of transfer. Imagine a set of distributed Soar knowledge engines. What would happen if they could exchange chunks? In other words, what would problem solving look like if a set of affiliated and distributed Soar problem solving engines could directly exchange their chunks – their intimate knowledge of a task? Base Task The task selected was the 8-puzzle (see Figure 1). In this task, there is a 3 x 3 matrix of eight randomly assigned numbers (the ninth being a space) and the problem is to find the series of operators (moves) that shuffle the set into a properly ascending sequence, by moving the tiles, oneby-one, to the open space. This task has 9! initial states and, depending on how you define it, can generate an uninformed search expansion of about 3.4 x 10. 1 Also called the N x N sliding tile problem, often appearing as a hand-held game where tiles are adjacently slid and interchanged until the right sequence/pattern is achieved. 318 FLAIRS 2002 From: FLAIRS-02 Proceedings. Copyright © 2002, AAAI (www.aaai.org). All rights reserved.