Application of machine learning to the maintenance of knowledge-based performance

Abstract

Integration of machine learning methods into knowledge-based systems requires greater control over the application of the learning methods. Recent research in machine learning has shown that isolated and unconstrained application of learning methods can eventually degrade performance. This paper presents an approach called performance-driven knowledge transformation for controlling the application of learning methods. The primary guidance for the control is performance of the knowledge base. The approach is implemented in the PEAK system. Two experiments with PEAK illustrate how the knowledge base is transformed using different learning methods to maintain performance goals. Results demonstrate the ability of performance-driven knowledge transformation to control the application of learning methods and maintain knowledge base performance.