Techniques for integrating machine learning with knowledge acquisition

Knowledge acquisition is frequently cited as the greatest bottleneck in the development of expert systems. Two primary approaches to knowledge acquisition are elicitation of knowledge from experts (traditional knowledge acquisition) and machine learning. Both approaches have strengths and weaknesses. Experts can draw upon practical experience and both domain specific and general knowledge. In addition, they often have access to well established procedures relating to the target domain. Elicitation of knowledge from experts is limited, however, by the difficulty of articulating knowledge; reluctance in some circumstances to share knowledge; preconceptions and biases that might inappropriately influence an expert; and the limits of an expert’s knowledge. In contrast, machine learning systems provide a capacity to infer procedures from examples; can perform extensive logical analysis; and are not subject to the same types of preconceptions and biases as an expert. They are hampered, however, by limited access to general and domain-specific knowledge and the difficulties of obtaining comprehensive example sets. Further, machine learning is only possible where much of the knowledge acquisition task has already been completed. Machine learning requires a description of the problem domain and collection of example cases from which to learn. In attribute-value machine learning, the domain description consists of a set of attributes and their allowable values along with a collection of class values. These, together with the formalism used for expressing the decision procedures, specify a space of possible solutions that the system might ‘learn’. The learning system then explores this space of solutions seeking one that best fits the training examples. If a suitable space of possible solutions is specified, learning is relatively straightforward. If a poor solution space is specified, effective learning is impossible. Thus, machine learning requires prior ontological analysis and the specification of a suitable class of models to explore.