Overview of statistically hedged prediction methods: From off-line to real-time data analysis

This work summarizes the latest results on prediction with newly developed estimators based on statistical significance. These predictors implement conformal predictions and have been applied to classification tasks for data of the TJ-II stellarator. In particular, different adaptations to solve a 5-class image classification problem for the TJ-II Thomson scattering (TS) are presented. Off-line (nearest neighbour and support vector machines based) and real-time (SVM based) versions of conformal predictors have been developed. In all cases, if the classifications are reliable, the predicted images are incorporated to the training dataset for future predictions. The nearest neighbour classifier (NNC) obtains a success rate of 97% with confidence 0.96 and a mean credibility of 0.61. The CPU time to predict shows a linear dependence with the number of images in the training set (t = 0.519n + 100.212 s). The SVM classifiers are used in the one versus the rest approach. The off-line version provides a success rate of 99%, a confidence of 0.99 and an average credibility of 0.55. The CPU time also follows a linear law with the number of images in the training set (t = 15.023 × 10−3n + 4.523 s). The real-time classifier achieves a success rate of 96% and a mean confidence and credibility of 0.99 and 0.53, respectively. In this case, after 395 classifications, the CPU time per image to classify remains constant: 89.7 ± 14.1 ms.

► Off-line and real-time multi-class image classification applications (conformal prediction based) have been developed.
► We compare two different implementations (nearest neighbour and SVM) in the off-line applications.
► A real-time predictor (SVM based) takes a constant time of 89.7 ± 14.1 ms to classify each image.
► The one-versus-the-rest approach is used in the SVM multi-class classifiers.