Efficient global path planning remains a fundamental requirement in mobile robot navigation, especially in large and cluttered environments where conventional methods often suffer from slow convergence, obstacle-edge contact, path overlap, and irregular zigzag trajectories. To address these limitations, this paper proposes a hybrid path-planning framework, called DAGA, that combines Deep Q-Network (DQN)-guided Ant Colony Optimization (ACO) with Genetic Algorithm (GA) refinement. In the proposed framework, the DQN-guided ACO stage generates safe feasible paths by improving state-transition decisions during exploration, while the GA stage refines these candidate paths to improve smoothness and reduce path length. An early stopping mechanism and an adaptive path-selection strategy are further introduced to reduce unnecessary computation and preserve useful population diversity during refinement. The proposed method was evaluated on three grid environments of increasing complexity, namely 15  30, and 60 × 60, and was further validated on two 100 × 100 real-world geographical maps. Experimental results over repeated runs show that DAGA consistently produces shorter, safer, and smoother paths than conventional ACO and also improves upon the intermediate DA stage. In the most complex 60 × 60 map, DAGA reduced the mean path length from 126.24 to 82.07, improved the smoothness cost from 19.355 to 5.003, and completely eliminated obstacle-edge touch and path overlap across all reported runs. Relative to ACO, this corresponds to approximately 34.99% shorter paths together with full removal of unsafe boundary contact and overlap. A broader 30-run baseline comparison on the same large map further shows that DAGA achieves the shortest path among the reported baselines and is the only method with 100% success, 0% obstacle-edge touch, and 0% path overlap. These findings indicate that DAGA provides an effective and reliable solution for global mobile robot path planning in complex large-scale environments.