Knowledge Organization and Management

Knowledge Organization and Management

The advantage of using structured knowledge representation schemes (frames, associative networks, or object-oriented structures) over unstructured ones (rules or FOPL clauses) should be understood and appreciated at this point. Structured schemes group or link small related chunks of knowledge together as a unit. This simplifies the processing operations, since knowledge required for a given task is usually contained within a limited semantic region, which can be accessed as a unit or traced through a few linkages.

But, as suggested earlier, representation is not the only factor, which affects efficient manipulation. A program must first locate and retrieve the appropriate knowledge in an efficient manner whenever it is needed. One of the most direct methods for finding the appropriate knowledge is exhaustive search or the enumerations of all items in memory. This is also one of the least efficient access methods. More efficient retrieval is accomplished through some form of indexing or grouping. We consider some of these processes in the next section where we review traditional access and retrieval methods used in memory organizations. This is followed by a description of less commonly used forms of indexing.

A “smart” expert system can be expected to have thousands or even tens of thousands of rules (or their equivalent) in its KB. A good example is XCON (or RI), an expert system which was developed for the Digital Equipment Corporation to configure their customer’s computer systems. XCON has a rapidly growing KB, which, at the present time, consists of more than 12,000 production rules. Large numbers of rules are needed in systems like this, which deal with complex reasoning tasks. System configuration becomes very complex when the number of components and corresponding parameters is large (several hundred). If each rule contained above four or five conditions in its antecedent or If part and an exhaustive search was used, as many as 40,000-50,000 tests could be required on each recognition cycle. Clearly, the time required to perform this number of tests is intolerable. Instead, some form of memory management is needed. We saw one way this problem was solved using a form of indexing with the RETE algorithm described in the preceding chapter, More direct memory organization approaches to this problem are considered in this chapter.

We humans live in a dynamic, continually changing environment. To cope with this change, our memories exhibit some rather remarkable properties. We are able to adapt to varied changes in the environment and still improve our performance. This is because our memory system is continuously adapting through a reorganization process. New knowledge is continually being added to our memories, existing knowledge is continually being revised, and less important knowledge is gradually being forgotten. Our memories are continually being reorganized to expand our recall and reasoning abilities. This process leads to improved memory performance throughout most of our lives.

When developing computer memories for intelligent systems, we may gain some useful insight by learning what we can from human memory systems. We would expect computer memory systems to possess some of the same features. For example, human memories tend to be limitless in capacity, and they provide a uniform grade of recall service, independent of the amount of information store. For later use, we have summarized these and other desirable characteristics that we feel an effective computer memory organization system should possess.

It should be possible to add and integrate new knowledge in memory as needed without concern for limitations in size.

Any organizational scheme chosen should facilitate the remembering process. Thus, it should be possible to locate any stored item of knowledge efficiently from its content alone.

The addition of more knowledge to memory should have no adverse effects on the accessibility of items already stored there. Thus, the search time should not increase appreciably with the amount of information stored.

The organization scheme should facilitate the recognition of similar items of knowledge. This is essential for reasoning and learning functions. It suggests that existing knowledge be used to determine the location and manner in which new knowledge is integrated into memory.

The organization should facilitate the process of consolidating recurrent incidents or episodes and “forgetting” knowledge when it is no longer valid or no longer needed.

These characteristics suggest that memory be organized around conceptual clusters of knowledge. Related clusters should be grouped and stored in close proximity to each other and be linked to similar concepts through associative relations. Access to any given cluster should be possible through either direct or indirect links such as concept pointers indexed by meaning. Index keys with synonomous meanings should provide links to the same knowledge clusters. These notions are illustrated graphically in Fig 9.1 where the clusters represent arbitrary groups closely related knowledge such as objects and their properties or basic conceptual categories. The links connecting the clusters are two-way pointers which provide relational associations between the clusters they connect.