Hierarchical Dirichlet Process mixture models (HDPMM) have recently been introduced not only in cognitive psychology but also in cognitive robotics. It was postulated in the context of categorization research that HDPMMs not only combine the strengths of all previous rational categorization procedures, but also unify the two prominent theories of categorization, the exemplar and prototype view, in a common categorization model. Consequently, findings were successfully modeled, for which the striking properties of an HDPMM, the dynamic complexity adaptation of the model to existing data, and the ability of cluster sharing appeared to be important key mechanisms. Researchers from cognitive robotics interpreted these successes as evidence for a good model of human categorization. In fact, HDPMMs are, however, a group of rational models that have so far hardly been explored. Hence, in this thesis the HDPMM introduced in cognitive psychology by Griffiths, Canini, Sanborn and Navarro (2007) is extensively evaluated. The study investigates the predictive power of the model regarding seven classical findings on human categorization, which are serious challenges for many of the current categorization models. It is demonstrated that two modifications to the model of Griffiths et al. (2007) are sufficient to successfully predict the majority of the findings. Furthermore, the HDPMM is compared with a prominent and well-established categorization model, the Supervised and Unsupervised Stratified Adaptive Incremental Network (SUSTAIN) with regard to data fitting and model flexibility, in which the modified model of Griffiths and colleagues can assert itself in the majority of the experiments. Finally, further possibilities for improvement, especially with regard to the application in the robotics context, are discussed.    «

Hierarchical Dirichlet Process mixture models (HDPMM) have recently been introduced not only in cognitive psychology but also in cognitive robotics. It was postulated in the context of categorization research that HDPMMs not only combine the strengths of all previous rational categorization procedures, but also unify the two prominent theories of categorization, the exemplar and prototype view, in a common categorization model. Consequently, findings were successfully modeled, for which the stri...    »