PULSE: A Pulsar Searching Model with Genetic Algorithm Implementation for Best Pipeline Selection and Hyperparameters Optimization

Abstract

Pulsars enabled astronomers to study neutron stars and verify general relativity under intense gravitational field conditions. However, finding pulsars is not as easy as it seems because most of them have weak pulses that get drowned in the background noise and hence do not get detected. This paper presents a novel way of classifying radio emission patterns collected from a radio telescope whether it is from a pulsar or not through machine learning and genetic algorithm. The dataset was acquired from the High Time Resolution Universe (HTRU) survey two which contains eight numerical features and one target variable describing the pulse profile. Synthetic Minority Oversampling Technique (SMOTE) was applied to the dataset to fix the imbalance between classes. A genetic algorithm library was used to automatically select the best feature preprocessing method, feature selection/reduction technique, machine learning model inside the scikit-learn library, and hyperparameter settings. The genetic algorithm suggested using a single stack and multiple stack classifiers for different sets of features. The optimum level of hyperparameters was also given with the help of the same algorithm. The selected pipelines consistently reported a score of more than 99% in all the evaluation metrics used.