Hybrid Metaheuristic-Q-Learning Framework for Robust Feature Selection in Cyber Intrusion Detection Systems

Abstract

Selection of features from high-dimensional network traffic data has been a key problem that hinders an efficient intrusion detection system, because unnecessary features reduce computational power as well as classification rates. In this paper, a hybrid optimization is suggested that methodically incorporates the methods of population-based metaheuristic calculation with tabular Q-Learning to address wrapper-based feature selection on network intrusion datasets. Pure metaheuristic algorithms enable global exploration by using diverse search strategies based on natural behaviors, whereas an agent based on an integrated Q-Learning mechanisms gets to learn how to deploy context sensitive feature modification operators such as selective addition, removal, swapping and crossover based on the observed fitness changes during optimization. Three categories are rigorously compared: standalone metaheuristic algorithms, Q-Learning based selection approaches, and hybrid methods where Q-Learning improves elite solutions each iteration. The framework uses an fitness criterion, which balances accuracy of classification and feature subset cardinality, evaluated through multiple ensemble classifiers. The hybrid approaches demonstrate better convergence behavior and consistently outperform both pure metaheuristics and standalone reinforcement learning, according to extensive experimental analysis. This establishes an efficient method for dimensionality reduction in cybersecurity applications.