Basic expert system terminology is reviewed, and several uncertainty management paradigms are surveyed. The focus is on subjective probability theory, Dempster-Shafer theory, and possibility theory, although a number of other methods are mentioned. The benefits and limitations of the various schemes are examined, examples of expert systems within each school are presented, and some relevant open problems are discussed. >

Uncertainty management in expert systems

In expert systems, hierarchical reasoning can provide better accuracy and understandability. Here, we develop a method of learning hierarchical knowledge from a case library, in which each training instance is described by low level features and high level concepts (e.g., manifestations and diseases) but not by intermediate concepts (e.g., disease states). Learning intermediate knowledge involves exploiting the old partial intermediate knowledge or creating new intermediate concepts by observing the relationship between the low level features and high level concepts. Experiments in the domain of diagnosing causes of jaundice validate the method.

/pdf/learning-intermediate-concepts-in-constructing-a-2sgzuklp5y.pdf

Learning intermediate concepts in constructing a hierarchical knowledge base

In this paper we analyse how noise can affect Knowledge Acquisition from a Machine Learning perspective We present some methods to detect and treat noise that goes beyond modulating numerical coefficients and show that noise cannot be viewed as a single entity There are several different types of noise and noise is not only wrong information

/pdf/noise-and-knowledge-acquisition-53wacsnn5l.pdf

Noise and knowledge acquisition

Learning by induction can require a large number of training examples. We show the power of using a simulator to generate training data and test data in learning rules for an expert system. The induction program is RL, a simplified version of Meta-DENDRAL. The expert system is ABLE, a rule-based system that identifies and locates errors in particle beam lines used in high energy physics. A simulator of beam lines allowed forming and testing rules on sufficient numbers of cases that ABLE's performance is demonstrably accurate and precise.

Simulation-assisted inductive learning

A novel technique that applies the neural-network learning strategy of back-propagation to recognize semantically incorrect rules is presented. When the rule strengths of most rules are semantically correct, semantically incorrect rules can be recognized if their strengths are weakened or change signs after training with correct samples. In each training cycle, the discrepancies in the belief values of goal hypotheses are propagated backward and the strengths of rules responsible for such discrepancies are modified appropriately. A function called consistent-shift is defined for measuring the shift of a rule strength in the direction consistent with the strength assigned before training and is a critical component of this technique. The viability of this technique has been demonstrated in a practical domain.

Recognition of semantically incorrect rules: a neural-network approach

A high performance expert system can be built by exploiting machine learning techniques. A learning model has been designed and implemented that is capable of constructing a knowledge base, in the form of rules, from a case library and continuously updating it to accommodate new facts. This model is designed primarily for EMYCIN-like systems in which there is uncertainty about data as well as about the strength of inference and in which the rules chain together to infer complex hypotheses. These features greatly complicate the learning problem. In machine learning, two issues that cannot be overlooked practically are efficiency and noise. A subprogram, called "CONDENSER", is designed to remove irrelevant features during learning and improve the efficiency. The noise can be handled by optimizing the result to achieve minimal prediction errors.
Another subprogram has been developed to learn meta-level rules which guide the invocation of object-level rules and thus enhance the performance of the expert system using the object-level rules.
Using the ideas developed in this work, an expert program called JAUNDICE has been built, which can diagnose the likely disease and mechanisms of a patient with jaundice. Experiments with JAUNDICE show the developed theory and method of learning are effective in a complex and noisy environment where data may be inconsistent, incomplete, and erroneous.

Li-Min Fu

Papers

Learning object-level and meta-level knowledge in expert systems (machine learning)