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 White et al. (2000). Seven input parameters and one output, evaluated to 1 for quasars and 0 for nonquasars, 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% at 80% completeness. For comparison the quasar fraction from the original candidate list was 56%. The accuracy is similar to that found by White et al. using supervised learning with oblique decision trees. This performance probably approaches the maximum value achievable with this database. Probabilities for the 98 candidates without spectroscopic classification in White et al. are presented and compared with the results from their work, showing a 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 databases.