A Guided Hybrid Genetic Algorithm for Feature Selection with Expensive Cost Functions

We present a guided hybrid genetic algorithm for feature selection which is tailored to minimize the number of cost function evaluations. Guided variable elimination is used to make the stochastic backward search of the genetic algorithm much more efficient. Guiding means that a promising feature set is selected from a population and suggestions (for example by a trained Random Forest) are made which variable could be removed. It uses implicit diversity management and is able to return multiple optimal solutions if present, which might be important for interpreting the results. It uses a dynamic cost function that avoids prescribing an expected upper limit of performance or the number of features of the optimal solution. We illustrate the performance of the algorithm on artificial data, and show that the algorithm provides accurate results and is very efficient in minimizing the number of cost function evaluations.We present a guided hybrid genetic algorithm for feature selection which is tailored to minimize the number of cost function evaluations. Guided variable elimination is used to make the stochastic backward search of the genetic algorithm much more efficient. Guiding means that a promising feature set is selected from a population and suggestions (for example by a trained Random Forest) are made which variable could be removed. It uses implicit diversity management and is able to return multiple optimal solutions if present, which might be important for interpreting the results. It uses a dynamic cost function that avoids prescribing an expected upper limit of performance or the number of features of the optimal solution. We illustrate the performance of the algorithm on artificial data, and show that the algorithm provides accurate results and is very efficient in minimizing the number of cost function evaluations