Machine learning methods have been used in cosmological studies to estimate variables that would be hard or costly to measure precisely, like, for example, estimating redshifts from photometric data. Previous work showed good results for estimating photometric redshifts using nonlinear regression based on an artificial neural network (Multilayer Perceptron). One of the problems identified on the previous work is that a single network with multiple neurons on the hidden layer for the regression may not be able to yield the best results. In this work we explore a hybrid neural network approach that uses a Self-Organizing Map to separate the original data into different groups, then applying the Multilayer Perceptron to each neuron on the Self-Organizing Map to obtain different regression models for each group. Preliminary results indicate that in some cases better results can be achieved, although the computational cost may be increased.
Link to PDF (may not be available yet): P1-11.pdf