Self-tune linear adaptive-genetic algorithm for feature selection

Abstract

Genetic algorithm (GA) is an established machine learning technique used for heuristic optimisation purposes. However, this natural selection-based technique is prone to premature convergence, especially of the local optimum event. The presence of stagnant performance is due to low population diversity and fixed genetic operator setting. Therefore, an adaptive algorithm, the Self-Tune Linear Adaptive-GA (STLA-GA), is presented in order to avoid suboptimal solutions in feature selection case studies. STLA-GA performs parameter tuning for mutation probability rate, population size, maximum generation number and novel convergence threshold while simultaneously updating the stopping criteria by adopting an exploration-exploitation cycle. The exploration-exploitation cycle embedded in STLA-GA is a function of the latest classifier performance. Compared to standard feature selection practice, the proposed STLA-GA delivers multi-fold benefits, including overcoming local optimum solutions, yielding higher feature subset reduction rates, removing manual parameter tuning, eliminating premature convergence and preventing excessive computational cost, which is due to unstable parameter tuning feedback.