基于强化学习的算力资源度量方法

Computational power resource measurement method based on reinforcement learning

工业边缘计算中,节点具有分布零散、异构以及算力受限的特点,为保障算力供给通常采用任务与平台紧耦合的模式,然而该模式易使生产系统刚性化,资源复用效率低,冗余资源成本高昂。针对这一问题本文提出了多维度任务分析与强化学习相结合的算力度量方法,首先对工业场景中计算任务的时、空复杂度,计算类型等特征进行细粒度分析,通过构建任务模型与计算模型分析各计算任务特征与资源需求比例之间的关系;随后,基于上述分析构建马尔科夫决策过程并把状态、动作、奖励建立为三元组,将奖励值定义为对任务执行时间的预测优化问题;最后,设计基于深度Q网络的计算任务算力度量方法,对不同形式的计算任务进行算力需求量化,并通过与设备实际算力消耗进行分析对比,验证所提方法可有效降低不必要的资源损耗。实验结果显示,所提出的模型和方法预测任务执行时间成功率可达99.37%,相较于Q-Learning等算法提升了约5.84%、7.54%和34.23%,可有效实现边缘侧的算力度量。

In industrial edge computing,nodes have the characteristics of scattered distribution,heterogeneity and limited computational power resource.In order to guarantee the supply of computational power resource,the mode of tight coupling between task and platform is ually adopted,bowever,this mode easily leads to the rigidity of production system,low resource reuse rate and high cost of redundant resources.In order to solve this problem,a computational power measurement method that combines multi-dimensional tack analysis and reinforcement leaning is proposed.Firstly,fine-grained analysis was carried out on the characteristics of computing tasks in industrial scenarios,such as time complexity,space complexity and computing type,and the relationship between the characteristics of each computing task and the proportion of resource requirements was analyzed h constructing taak model and computing moodel.Then Markov decision process was constructed based on the abowe analysis,and the state,action and reward were established as triples,and the reward value was defined as the prediction optimizaion problem of task execution time.Finally,a conmputational power measurement methbod based on DQN(Deep Q Netwok)is designed to quantify the computational power requirements of diferent forms of computing tasks.By analyzing and comparing with the actual computational power consumption of equipment,the proposed method can effectively reduce unnecessary resource consumption.Experimental results show that the success rate of the proposed model and method in predicting task execution time can reach 99.37%,which is about5.84%,7.54%,and 34.23%higher than algorithms such a Q-Leamning,and can efectively realie edge side computational power measurement.