Abstract: To optimize the performance， utilization and balance of physical servers in the process of virtual machine placement， an improved placement algorithm based on reinforcement learning is proposed. Firstly， the single reward in reinforcement learning algorithm is replaced by multi-dimensional rewards composed of dynamic objectives， and then the multi-dimensional reward values corresponding to each placement scheme are output by the predictor. The input of the predictor is the resource pool status， optimization target， and target proportion. Finally， the proportions of different optimization objectives are adjusted according to the resource pool status to optimize the targets. The results of CloudsimPy show that the improved algorithm can deploy more virtual machine requests with the higher placement success rate and lower delay consumption under the same resource.

Key words: virtual machine placement; deep reinforcement learning; resource utilization; load balance; virtual machine performance; K-mean clustering