面向深度增强学习的网络数据流优化研究

Research on Network Data Flow Optimization for Deep Reinforcement Learning

探讨面向分布式深度增强学习框架的网络数据流优化,设计使深度学习训练能容忍无损网络的数据流优化策略。对于参数服务器同步架构,假设计算节点和服务器之间的网络通信会产生丢包,该优化策略需要保证分布式训练在无损网络中的性能。采用理论分析与实验评估相结合的方式来验证所提出策略的性能。

This paper explores network data flow optimization for distributed deep enhancement learning frameworks,and designs deep learning training to tolerate data flow optimization strategies for lossless networks.For the parameter server synchronization architecture,it is assumed that network communication between the compute node and the server will result in packet loss.The optimization strategy of this paper needs to ensure the performance of distributed training in the lossless network.This uses a combination of theoretical analysis and experimental evaluation to verify the performance of the proposed strategy.