With the development of cloud gaming, Internet of Vehicle, virtual reality and industrial automation technologies, services with high bandwidth demands are emerging continuously. The rational allocation and efficient utilization of network bandwidth resources have become particularly important. However, traditional resource reservation is difficult to cope with dynamic network traffic. Inflexible bandwidth allocation and long-term bandwidth occupation can lead to the wastage of network resources. To address these issues, we propose an uneven bandwidth reservation strategy in software defined network (SDN) architecture. The proposed strategy aims to minimize the variance of link bandwidth utilization as the objective function and uses the Q-learning algorithm to iteratively obtain the optimal reserved bandwidth. Experiments show that the proposed strategy not only meets quality-of-service (QoS), but also effectively improves reserved bandwidth utilization. Furthermore, the proposed strategy reduces transmission time of traditional traffic and optimizes overall network performance.