A framework for the investigation of multiparametric dependences applied to total radiated power of JET plasmas

A framework is developed for investigating complex multivariate relationships in a dataset. This is based on using the universal approximation abilities of a multi-layer perceptron (MLP) neural network to predict a quantity of interest from a large set of parameters. A measure of redundancy is derived, and used in such a way that the average influence on the predicted quantity from any parameter can be estimated. Input parameters can be ordered in terms of increasing redundancy and therefore assist in finding the most important parameters a phenomenon of interest depends upon. In spite of the problem being multi-dimensional, the functional form of the one-to-one relationship between a parameter and a quantity of interest can be visualized. This framework is then used together with sensitivity analysis to investigate the dependence of the total radiated power of JET plasmas on a large number of parameters, leading to the identification of a much smaller set of parameters to be used in an effective MLP predictor of total radiated power.