Using A Priori Knowledge to Prestructure ANNs

The objective of the present work is to develop a constructive method that uses certain a priori information about a problem domain to design the starting structure of an artificial neural network (ANN). The method explored is based on a general systems theory methodology (here called GSM) that calculates a kind of structural information of the problem domain via analyzing I/O pairs from that domain. A modularized ANN structure is developed based on the GSM information provided. The notion of performance subset (PS) of an ANN structure is described, and extensive experiments on 3-input, 1-output Boolean mappings indicate that the resulting modularized-ANN design is 'conservative' in the sense that the PS of the modularized ANN contains at least all the mappings included in the GSM category used to design the ANN. Partial experiments on 5-input, 1-output Boolean functions indicate further success. The extended experimental results also suggest the possibility of using a measure of the learning curve of specified ANNs on a series of (in this case Boolean) functions to serve as a proxy measure for the complexity of those functions. This proxy measure seems to correlate well with a measure known as Boolean Length. Determining a function's Boolean Length is a non-trivial undertaking; perhaps it will turn out that training an ANN on the function and measuring its learning experience will be a useful measure of function complexity, and easier to determine than the function's Boolean Length.