Selection of quasar candidates from combined radio and optical surveys using neural networks

Abstract

The application of supervised artificial neural networks (ANNs) for quasar selection from combined radio and optical surveys with photometric and morphological data is investigated, using the list of candidates and their classification from the work of White et al. Seven input parameters and one output, evaluated to 1 for quasars and 0 for non-quasars during the training, were used, with architectures 7: 1 and 7: 2: 1. Both models were trained on samples of ∼800 sources and yielded similar performance on independent test samples, with reliability as large as 87 per cent at 80 per cent completeness (or 90 to 80 per cent for completeness from 70 to 90 per cent). For comparison, the quasar fraction from the original candidate list was 56 per cent. The accuracy is similar to that found by White et al. using supervised learning with oblique decision trees and training samples of similar size. In view of the large degree of overlapping between quasars and non-quasars in the parameter space, this performance probably approaches the maximum value achievable with this data base. Predictions of the probabilities for the 98 candidates without spectroscopic classification in White et al. are presented and compared with the results from their work. The values obtained for the two ANN models and the decision trees are found to be in good agreement. This is the first analysis of the performance of ANNs for the selection of quasars. Our work shows that ANNs provide a promising technique for the selection of specific object types in astronomical data bases.