Automatic Generation of Domain Specific Inference Program for Knowledge Processing System

Abstract

The generic task method is a powerful route to easier construction of knowledge processing systems. The problem solving process is regarded as an information processing task, and the knowledge processing system is constructed by adding task implementation knowledge to the generic task, which is a pre-defined building block of an expert tool for a specific information processing task. The weak point of the method is that the collection of problem solving methods is not systematic. According to the characteristics of a problem to solved, users select a best-fit generic task and adjust formalization of the problem solving method to this selected generic task. In most cases, the load of the adjusting process is not negligible. Systematic classification of problem solving methods is the key to enhancement of the expert system tool. In order to construct the knowledge processing system more easily, we propose an expert tool which has a problem-directed inference program generation function. With the proposed expert tool, users then construct a knowledge processing system in the following way. (1) Users divide the process into search modules. (2) Users formalize search module as a search, (3) Users input the search strategy by selecting the classification items of the search classification tree, which the tool has prepared. (4) The tool produce the problem-directed inference program by setting appropriate program parts in the template of the general search algorithm. The program parts are prepared as an abstract data type in the tool. (5) The tool adds task implementation knowledge to the problem-directed inference program, and completes the search module. (6) The tool integrates each search module and completes the knowledg processing system. The tool is applied to construction of three types of scheduling systems. They are a maintenance scheduling system, construction scheduling system, and job-shop scheduling system. By using the proposed tool, problem-directed inference programs are produced. Comparing the inference programs implemented by using a conventional tool, equivalent performance and a two-thirds reduction in program step numbers required as the programmers' input are realized.