<p>The cluster scheduler assigns jobs to available computing resources. The increasing complexity of the cluster systems, coupled with the high dynamic workload, imposes great burden on cluster resource scheduling. The article first employs a deep reinforcement learning algorithm to achieve an improved job scheduling strategy, which uses convolutional neural networks to capture real-time state information of virtual machines in data centers. Subsequently, according to the set expert strategy, imitation learning is used to reduce the number of exploration steps of reinforcement learning, so as to shorten the training time of the optimal strategy. Finally, different reward functions are set according to different scheduling objectives, and the optimization strategy is solved by deep reinforcement learning strategy. Based on the job scheduling objectives, job scheduling model learns the improvement policy by interacting with the scheduling environment, and adapts it dynamically in response to workload variations. This effectively solves the issue of ineffective job scheduling caused by varying job types and sizes, and the dynamic change of virtual machine state. The job scheduling algorithm and other benchmark scheduling algorithms are evaluated through virtualized experiments, the results indicate that the job scheduling scheme decreases both the average turnaround time and the average weighted turnaround time. Compared with the current state-of-the-art scheduling algorithm DeepRM, the proposed scheduling algorithm is 37.5% faster in terms of the convergence speed. Moreover, the proposed scheduling algorithm reduces the average weighted turnaround time and the average turnaround time by 4.07% and 6.99%, respectively.</p>