Adaptive deterministic annealing for two applications: competing SVR of switching dynamics and travelling salesman problems

A deterministic annealing approach has been proposed to clustering by Rose et al. (1990; 1992) based on the maximum entropy principle. They obtain the association probabilities at a given average variance. The corresponding Lagrange multiplier is inversely related to the "temperature" and is used to control the annealing process. We propose an adaptive annealing schedule where the Lagrange multiplier is treated as an unknown parameter and is estimated by an expectation-maximization step. This technique is applied to using support vector regression (SVR) or switching dynamics. We also give some preliminary results on traveling salesman problems (TSP).