Knowledge Extraction from the Neural ‘Black Box’ in Ecological Monitoring

Phytoplankton biomass within the Saginaw Bay ecosystem (Lake Huron, Michigan, USA) was characterized as a function of select physical/chemical indicators. The complexity and variability of ecological systems typically make it difficult to model the influences of anthropogenic stressors and/or natural disturbances. Here, Artificial Neural Networks (ANNs) were developed to model chlorophyll a concentrations, a measure for water-column phytoplankton biomass and a proxy for system-level health. ANNs act like “black boxes” in the sense that relationships are encoded as weight vectors within the trained network and as such, cannot easily support the generation of scientific hypotheses unless these relationships can be explained in a comprehensible form. Accordingly, the ‘knowledge’ and/or rule-based information embedded within ANNs needs to be extracted and expressed as a set of comprehensible ‘rules’. Such extracted information would enhance the delineation and understanding of ecological complexity and aid in developing usable prediction tools. Comparisons of various computational approaches (including TREPAN, an algorithm for constructing decision trees from neural networks) used in extracting rule-based information from trained Saginaw Bay ANNs are discussed.