Motor Imagery–based Brain–Computer Interface (MI BCI) systems rely heavily on the extraction and selection of discriminative features from Electroencephalography (EEG) signals. However, EEG data are inherently noisy, high-dimensional, and non-stationary, making feature selection a challenging optimization problem. This paper introduces a novel reinforcement learning–driven optimization framework called the World Hyper Heuristic (WHH) algorithm. Unlike traditional metaheuristics that depend on fixed operators, WHH dynamically selects among multiple Low-Level Heuristics (LLHs) based on environmental feedback, enabling a more effective balance between exploration and exploitation. The algorithm is evaluated on benchmark MI datasets from the BCI Competition series and compared with five widely used optimization algorithms: Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Differential Evolution (DE), Grey Wolf Optimizer (GWO), and Random Search (RS). Results demonstrate that WHH consistently outperforms all baselines in classification accuracy, F1 score, kappa coefficient, feature reduction rate, convergence speed, and cross-session stability. WHH achieves up to 12.4% higher accuracy, reduces feature dimensionality by 43–57%, and improves stability by 18% compared to the best competing method. These findings highlight the potential of reinforcement learning–based hyperheuristics as a powerful and adaptive optimization paradigm for next-generation BCI systems.