Date of Award


Document Type


Degree Name

Master of Science


Department of Electrical and Computer Engineering

First Advisor

Steven K. Rogers, PhD


Recent advances in machine learning theory have opened the door for applications to many difficult problem domains. One area that has achieved great success for stock market analysis/prediction is artificial neural networks. However, knowledge embedded in the neural network is not easily translated into symbolic form. Recent research, exploring the viability of merging artificial neural networks with traditional rule-based expert systems, has achieved limited success. In particular, extracting production (IF.. THEN) rules from a trained neural net based on connection weights provides a valid set of rules only when neuron outputs are close to 0 or 1 (e.g. the output sigmoid function is saturated). This thesis presents two new ways to interpret neural network knowledge. The first, called Knowledge Math, extends the use of connection weights, generating rules for general (i.e. non-binary) input and output values. The second method, based on decision boundaries, utilizes the inherent border between output classification regions to draw symbolic interpretation. The Decision Boundary method generates more complex symbolic rules than Knowledge Math, but provides valid feature relationships in the uncertain regions around the midpoints of the neuron output functions. The main result is a complementary relationship between Knowledge Math and Decision Boundaries, as well as subsymbolic and symbolic knowledge representations for a general multi-layer perceptron.

AFIT Designator


DTIC Accession Number