29th Symposium on Fusion Technology (SOFT 2016)

P4.063 Boosting learning for robust classification of TJ-II nuclear fusion databases

8 Sep 2016, 14:20

1h 40m

Foyer 2A (2nd floor), 3A (3rd floor) (Prague Congress Centre)

Board: 63

Poster D. Diagnostics, Data Acquisition and Remote Participation P4 Poster session

Speaker

Gonzalo Farias (Escuela de Ingenieria Electrica)

Description

Huge databases are a common situation in fusion. Physical properties of plasma are studied by thousands of signals, sampled at very high frequencies, producing enormous amount of data. A medium-size nuclear fusion device such as TJ-II can generate discharges that last around 500 milliseconds, reaching up to 100 Mbytes per one simple shot. Larger fusion devices such as JET can produce 10Gbytes per discharge, and ITER could storage 1Tbytes per a 30 minutes shot. The thousands acquired signals involve the analysis of data in high-dimensional spaces. In such spaces, the data become sparse, which makes difficult the searching of patterns with similar properties, reducing the efficiency and increasing the overfitting of learning algorithms. This issue, which is known in the literature as the course of dimensionality, can be faced by using suitable feature extraction methods to reduce the input space into a low-dimensional space. However the selection of feature reduction techniques is not straightforward, and commonly it is a time consuming task that requires an important effort. During last years the use of boosting algorithms is become very popular to avoid overffiing and to obtain generalized classifiers in problems with high-dimensional spaces. Boosting is an approach to machine learning to achieve a highly accurate and robust classification by combining many relatively weak and simple rules. The AdaBoost algorithm was the first practical boosting algorithm, and is one of the most widely studied, with applications in several fields.  This article describes the use of AdaBoost for building robust classifiers of patterns in fusion databases. In order to show the benefits of the approach, images from the Thomson Scattering diagnostic, and time-domain signals of the TJ-II database have been tested. The work includes a comparative study with previous results of other classifiers built with support vector machines and artificial neural networks.