The symbolic-neural method for creating models and control behaviors from examples

It was shown that neural networks train faster, with a higher probability of rule extraction if the training data is "direct" or "strictly monotonic," and traditional symbolic/algorithmic reasoning tools can be used to work with the intermediate data.

Abstract

A method of modelling extremely complex systems is proposed using back-propagation neural networks a part of a knowledge base. This modelling approach is particularly useful for the development of trained models of real physical systems as built for diagnosis of faults, or for the modelling of human behavior which can then be used to control systems. For modelling complex analog behaviors, back-propagation provides an approach to curve fitting and fuzzy logic because of its property of extraction. It was shown that neural networks train faster, with a higher probability of rule extraction if the training data is direct or Large networks can be broken into architectures of smaller networks a way of making data relationships strictly monotonic. Furthermore, rule-injection hints can be used to make a network learn nonmonotonic relationships faster with a higher probability of rule extraction. What makes an effective hint can be defined mathematically and generally makes intuitive sense. Dividing larger networks into architectures of smaller networks adds to the amount of symbolic data required to build systems, but this additional intermediate data then becomes available available in human intelligible (symbolic) form. Thus traditional symbolic/algorithmic reasoning tools can be used to work with the intermediate data. These theories were confirmed in a series of simple experiments involving some strictly monotonic training sets and an augmented XOR function. Finally, an experiment in modelling human control behavior of a planetary lander was built. This model also confirmed the findings on monotonic training data and the use of hints. Furthermore it provided an example of building models in varying stages of detail, all of which functioned to some degree, but the more detailed models performed the task more accurately and were trained in less time than their less detailed counterparts.