A Learning Classifier Scheme forDiscrete Variable Pattern Recognition Problems
I. K. Sethiandb · 1978
Theproblem ofclassifier design with discrete measure- mentvariables isconsidered fromtheswitching algebra point ofview. Theconcept oflearning isthenintroduced tospeed uptheclassifier design process. I.INTRODUCTION Theproblem ofdiscrimination using discrete measurement datahasbeenwellstudied (1)-(3). Inconnection withpattern recognition problems involving suchmeasurement variables, a classifier design technique wasproposed byStoffel (4) under the framework ofprime eventtheory. Inthis, eachdiscrete, multi- valued d-dimensional measurement vector istreated asaneventin themeasurement space. Associated witheach eventisacapability tocover other events. Thusaneventcanbeviewed topartition thesetofpossible measurement vectors (patterns) into twosub- sets. Onesubset comprises themeasurement vectors which are covered bytheevent, andtheother subset contains allthemea- surement vectors notcovered bytheevent. Utilizing this covering property, aclassifier canbedesigned intermsofasetofrepresent- ative events foreachcategory. Theclassification rule thenisto assign thegiven measurement vector thecategory whosesetof representative events covers thegiven measurement vector. Stoffel termssuchasetofrepresentative events asprime eventsand proposes twoalgorithms forfinding themfromagiven setof labeled samples. Inthis correspondence, wepresent alearning scheme forthe classifier based ontheaboveconcept. Itisassumed that atthe input totheclassifier, a pattern P isrepresented bya d- dimensional pattern vectorV(P), where