Frank, >>Also my measure would include learnability as a factor. It would not >>be driven purely by architectural demands and niceties. > >'Learnability'. Care to quantify that? :-)
This is where the classification work gets interesting. Quoting from Estes, page 40: "It has been demonstrated that some behavioral measures, eg the order in which individuals list exemplars of categories, are correlated with typicality ratings, one might wonder whether relative typicalities computed from similarity-to-category values could be interpreted as predicted probabilities of categorization responses." I may have misunderstood some parts of this work. But it looks as if various studies have confirmed the predictive power of this proposal. So learnability can be computed using similarity-to-category. The number that pops out the end is the probability that a subject will categorize an instance as belonging to the given category. There are various complications, such as working out values for similarity and what is called psychological distance. In some cases no similarity values need be known. For instance Estes gives the example of two different ways of classifying black/white triangles/squares. In one case the response probability is 1/(1+s) and in the other (1 + s.s)/((1+s)(1+s)). With s varying between 0 and 1. For all values of s the first response probability is always greater than or equal to the second. Thus, the first classification has the higher probability of subjects correctly classifying instances that belong to it. derek -- Derek M Jones tel: +44 (0) 1252 520 667 Knowledge Software Ltd mailto:[EMAIL PROTECTED] Applications Standards Conformance Testing http://www.knosof.co.uk - Automatic footer for [EMAIL PROTECTED] ---------------------------------- To unsubscribe from this list, mail [EMAIL PROTECTED] unsubscribe discuss To join the announcements list, mail [EMAIL PROTECTED] subscribe announce To receive a help file, mail [EMAIL PROTECTED] help This list is archived at http://www.mail-archive.com/discuss%40ppig.org/ If you have any problems or questions, please mail [EMAIL PROTECTED]
