NONMETRIC CONCEPTUAL CLUSTERING IN ECOLOGY AND ECOTOXICOLOGY

Ecological studies and multispecies ecotoxicological tests are based o n the examination of a variety of physical, chemical, and biological d ata with the intent of finding patterns in their changing relationship s over time. The data sets resulting from such studies are often noisy , incomplete, and difficult to envision. We have developed machine lea rning and visualization software to aid in the analysis, modeling, and understanding of such systems, and have applied it to the analysis of lake and stream field studies and aquatic microcosm toxicological tes ts. The software is based on nonmetric conceptual clustering, which at tempts to group the data into clusters that are strongly associated wi th several measured parameters. In each case, our tools not only confi rmed suspected ecological patterns, but also revealed aspects of the d ata that were unnoticed by ecologists using conventional statistical t echniques. Machine learning tools should, accordingly, become a standa rd part of the ecologist's armamentarium.