In traditional methods about DS/AHP, decision making results are poor effective resulting from information inference modes lacking of flexibilities. Therefore, the modeling procedures and the existing drawbacks in traditional methods are analyzed, and the flexible knowledge matrix is introduced to accommodate multiple inference information in terms of three relative inference types, i.e., part to whole, part to part, whole to part. After that, the theory model, the calculation model, and their equivalence theorem are proposed to recognize the optimal basic probability assignment(BPA) functions based on optimization principles. Finally, a numerical comparison analysis is applied to illustrate the presented method to be scientific and feasible.