The polyethylene industry faces persistent challenges in optimizing feedstock and energy consumption due to the highly nonlinear, non-closed-form behavior of polymerization reactors, which renders classical gradient-based methods ineffective and trial-and-error approaches costly. Although machine learning and metaheuristic algorithms have been explored, existing solutions suffer from three major limitations: inability to generate novel feedstock blends beyond training data, lack of dynamic parameter adaptation, and absence of closed-loop feedback between prediction and optimization. To address these gaps, this study proposes a hybrid framework that synergistically integrates a Genetic Algorithm (GA), a Reinforcement Learning (RL) agent, and a Gradient Boosting Machine (GBM). In this framework, the GA generates candidate feedstock combinations, the GBM—trained on 360 daily industrial records—predicts polyethylene and wax yields with 99.9% accuracy, and a Q-learning-based RL agent dynamically tunes three critical GA parameters based on real-time performance feedback, ensuring adaptive exploration and preventing premature convergence. Applied to actual data from a high-density polyethylene (HDPE) plant, the proposed framework increased average daily polyethylene output from 950 to 1,018 tons (a 7.16% improvement) and reduced daily wax generation from 12 to 7 tons (a 41.66% reduction). Statistical analysis across 30 independent runs, verified by the Wilcoxon signed-rank test (p < 0.05), confirms that these improvements are significant and consistently outperform five well-known metaheuristic algorithms (GA, PSO, GWO, DE, and ABC). The proposed RL-GA-GBM framework provides a robust, generalizable, and computationally efficient decision-support tool for petrochemical process optimization, effectively bridging the gap between theoretical research and industrial deployment.