<p>This study presents a novel learn-to-ensemble algorithm that dynamically designs problem-specific metaheuristics via deep reinforcement learning. The algorithm dynamically suggests combinations of metaheuristics using a Deep Q-Network (DQN) agent, adapting them to the optimization problem’s characteristics. The learned combinations are then applied to larger-scale problem instances to assess their generalization capability. Moreover, if two problem instances share similar characteristics, a heuristic that performs well on one is likely to perform well on the other. This study eliminates the need for proposing new meta-heuristic metaphors or methods. Evaluations across 20 numerical optimization problems from the IEEE Congress on Evolutionary Computation (CEC) 2021 and 2022 and combinatorial problems in Industrial Engineering (IE) demonstrate significant performance gains over individual algorithms. Computational and statistical analyses across 30 independent runs confirm the method’s accuracy and robustness, demonstrating superior average solution accuracy and consistency compared to benchmarks across most test problems. The algorithm achieved superior mean results in 75% of the CEC problems. In IE problems, it outperformed its counterparts in the mean results for all cases of the Knapsack Problem, Traveling Salesman Problem, and Vehicle Routing Problem, and achieved 90% superiority in the Job Shop Scheduling Problem. Each problem was run 30 times to ensure the results were statistically significant, confirmed by the Wilcoxon test.</p>