Application of Questionnaire Theory to Pattern Recognition

Analogies between Pattern Recogn i t ion and St ati stacal Communication Theory are pointed out. With the relat ive entropy between the "ideal a t t r i ­ bute" and "heuristic at t r ibute" sets as an index of performance, optimum selection of possible heur i s t i c operators resuJ ts in a recognition deci­ sion in probabi l i t ty , with t he smallest possi ble average number of tests during a "sequential ana­ lysis"'. On the other hand, a computer gencrati on of quasi-optimal operators may be done based on nume­ r ical attr ibutes using the same entropy index of performance. Experimental results are presented : Optimal operator choice is demonstrated for recognition of handwri tten capital le t te r words and for vocoder speech recogniti on. Computer bui ldi ng of quasi-optimal operators Is shown for d ig i t ized images. They are l imited to f i r s t level l inear numerical operators. The probabil ity distr ibut ion of the events is not modifi ed by any observation knowledge , which then gives no information on X. Information theory provides a measure of this information obtained by the observation knowledge Among many possible presentations let us re­ cal l br ie f ly the "Questionnaire theory" approach, cf. Pi card (13) , Terrenoi re (15) ; Th.e average minimum number of binary ques­ tions to be asked to obtain the prohahil i ty one answer is given by