The article develops a unified approach for hard optimizationproblems involving data association, i.e. the assignment of elementsviewed as “data” {xi}, to one of a set ofclasses, (Cj), so as to minimize the resulting cost. Thediverse problems which fit this description include data clustering,statistical classifier design to minimize probability of error,piecewise regression, structured vector quantization, as well asoptimization problems in graph theory, e.g. graph partitioning. Whereasstandard descent-based methods are susceptible to finding poor localoptima of the cost, the suggested approach provides some potential foravoiding local optima, yet without the computational complexity ofstochastic annealing. The approach we develop is based on ideas frominformation theory and statistical physics, and builds on the work ofRose, Gurewitz, and Fox (see IEEE Trans. on Inform. Theory, vol.38,p.1249-58, 1992) for clustering and related problems
展开▼
