On the potential of ruled-based machine learning for disruption prediction on JET

In the last years, it has become apparent that detecting disruptions with sufficient anticipation time is an essential but not exclusive task of predictors. It is also important that the prediction is accompanied by appropriate qualifications of its reliability and it is formulated in mathematical terms appropriate for the task at hand (mitigation, avoidance, classification etc.). In this paper, a wide series of rule-based predictors, of the Classification and Regression Trees (CART) family, have been compared to assess their relative merits. An original refinement of the training, called noise-based ensembles, has allowed not only to obtain significantly better performance but also to increase the interpretability of the results. The final predictors can indeed be represented by a tree or a series of specific and clear rules. Such performance has been proved by analysing large databases of shots on JET with both the carbon wall and the ITER Like Wall. In terms of performance, the developed tools are therefore very competitive with other machine learning techniques, with the specificity of formulating the final models in terms of trees and simple rules.