In today’s expanding and diversified data networks, static QoS mechanisms fall short in dynamically satisfying varying application requirements. A DRL-based self-adaptive routing scheme in an SDN was proposed. The agent was trained and tested on a simulated SDN environment with 799,223 samples. For dynamic traffic engineering, optimal parameters such as latency, packet loss, and throughput were minimized. The implementation on the POX controller for a real-world deployment demonstrated significant reductions; latency was reduced by 9.3%, packet loss was reduced by 90.9%, and bandwidth utilization was enhanced by 2.7%. Comparison of validation results with Q-Learning demonstrated the superiority of the performance metrics, showing DQN / Q-learning in all performance metrics (packet loss: 90.9% - 61.0%, latency: 9.3%–4.38%, generalization capability: 51.1%–25%). The DQN was able to reconcile contradicting QoS targets and can be the foundation of adaptive SDN infrastructures able to optimize a network autonomously in real time.