A heuristic simulated annealing algorithm of learning possibility measures for multisource decision making

In multisource decision making, non-normalized possibility measures are used to gauge the credibilities of each subset of information sources in order to reduce the uncertainty introduced by the sources and minimize decision error. The credibility of individual information source, i.e., the possibility density, is objectively learned from subjective chosen training data. A learning algorithm of the possibility measures is constructed based on the ideas of simulated annealing and feasible direction. The algorithm is designed for large training data set. Two multifeature image segmentation experiments and a comparison of learning speed are presented.