MONITORING RELAXATION ALGORITHMS USING LABELING EVALUATIONS.

Relaxation algorithms employ an initial stochastic classification and a probabilistic model for deferring final decision. In such algorithms it is desirable to evaluate a classification based on an initial stochastic classification and a probabilistic model. Such evaluation can monitor the classification and ensure reasonable results. Specific classifications are evaluated, rather than evaluating the intermediate probabilistic classifications as was considered in previous work. Since relaxation algorithms are not guaranteed to converge to a reasonable solution, monitoring them is useful as a stopping criterion.