Application of the AutoClass Artificial Intelligence Program to Asteroidal Data

Document Type

Conference Proceeding

Publication Date


Publication Title

American Astronomical Society, Department of Planetary Sciences


As our digital databases grow, datasets become less tractable and investigating alternative analysis techniques such as artificial intelligence algorithms becomes more important. One such program, AutoClass, which was developed by NASA's Artificial Intelligence Branch, uses Bayesian classification theory to automatically choose the most probable classification distribution to describe a dataset. To investigate its usefulness to the Planetary community, we tested its ability to reproduce the taxonomic classes as defined by Tholen and Barucci (1989). We started our evaluation by entering all Tholen identified C, S, or X type Eight Color Asteroid Survey asteroids with a color difference error of less than +/- 0.05 magnitudes. Of these 406 asteroids, AutoClass was able to firmly classify 346 (85%), identifying the remaining 60 asteroids as belonging to more than one class. Of the 346 asteroids that AutoClass classified, all but 3 (<1%) were classified as they had been in the Tholen classification scheme. The three that were misclassified had color errors estimated to be greater than +/- 0.04 magnitudes (though several other asteroids with such errors were classified correctly). To further test AutoClass, we expanded our reach to include all taxonomic types in the ECAS data, and further to include the nine wavelengths used to create the Bus and Binzel taxonomic superclasses (2002), with similar results. The initial successes of AutoClass and its ability to scan large domains for natural classes, showcase its exciting potential as a new discovery tool for Planetary scientists.