A Possibilistic Approach to Target Classification
Albert G. Huizing, Frans C. A. Groen · Studies in fuzziness and soft computing · 2000
This chapter describes an alternative to the Bayesian approach to target classification that is based on possibility theory. A possibilistic classifier minimizes the maximum cost of the classification decision taking into account the a posteriori possibilities of the target classes given the measured target attributes. The advantage of a possibilistic classifier when compared with a Bayesian classifier is that it requires only an ordinal ranking of the costs associated with the classification decisions and the uncertainty about the target class. Owing to its qualitative character, a possibilistic classifier is less sensitive to inaccuracies in a priori knowledge than a Bayesian classifier at the expense of a degraded performance in situations where accurate a priori knowledge is available. This robustness of the possibilistic classifier to inaccuracies in a priori knowledge is demonstrated in a case study where an average cost criterion is used to compare the performance of a possibilistic and a Bayesian classifier. It is shown that when the characteristics of the measured target attributes deviate strongly from the expected characteristics, the possibilistic classifier provides a lower average cost than a Bayesian classifier. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.