Corey D. SCHOU / Earnest (Skip)LOHSE / Kandi TURLEY-AMES / Keith LOHSE / Dotty SAMMONS / Michael SMUIN

(Idaho State University, Informatics Research Institute, Pocatello, USA)

This paper updates information on SIGGI (a web based AI/Neural Network (NN) software tool), applied to automated classification of projectile point types from the Columbia Plateau and Northwestern Plains of North America. The SIGGI autoclassification system is an extensible virtual analyst that initially uses a set of rules derived from Lohse (1985). The system applies classic AI/NN techniques to analysis of image databases. The image database techniques store and retrieve data for both knowledge elicitation from experts and NN analysis.
Accuracy of the system depends on the introduction of new image data sets, and the evaluation of new decision models based on knowledge elicitation. As a result, SIGGI incorporates knowledge from human experts to set parameters for its training and testing on image data sets. Introduction of new datasets allows SIGGI to refine the decision rules it uses to classify projectiles. After exposure to large numbers of images, SIGGI can extract the outline of the point and suggest the types it resembles. The only external input the system needs is a relative size (small or large) and, to clarify broken specimens, an initial classification as a lanceolate or triangulate. The technical specifications of how SIGGI works has been reported elsewhere (Lohse et al. 2004, 2005). To improve the characteristics of the NN analysis, the SIGGI model requires additional information about decision criteria. Part one of this project incorporates knowledge elicitation protocols based on human experts that define salient characteristics of the artifact types. This is performed in a controlled environment for experts. Psychologists will evaluate this data to establish protocols for NN improvement and improved SIGGI models.These protocols allow continuous training of SIGGI. Results demonstrate SIGGI’s extensibility to other geographic domains as well as its extensibility to other artifact classes. SIGGI will eventually be used as a smart user interface to help archaeologists sort, analyze, and classify a wide range of digital data.

Keywords: database, classification, multivariate analysis, North America