Basic Principles of Annealing for Large Scale Non-Linear Optimization

摘要: Computational Annealing, a class of optimization heuristics that are inspired by statistical physics phase transitions has been demonstrated to be highly effective for large, non-linear combinatorial problems. In many applications in computer vision and pattern recognition one encounters objective functions with very large number discrete possibly additional continuous variables. Typical cases such problems clustering, grouping image segmentation or assignment motion stereo analysis object recognition. For this type problems, standard integer programming techniques not applicable resort fast, yet avoid exponential unfavorable local minima. A particularly powerful, generic algorithms is provided simulated deterministic annealing techniques. Simulated the Gibbs sampler discussed first present basic concepts; then, theory presented great detail relation continuation methods discussed.