CAJ | 학술논문

我们为分布式业务系统提出了一种基于模型与预测的接入控制方法.首先,使用部分可观Markov决策过程(POMDP)来为业务系统建模.接着,将服务分配策略融入系统参数并且使用随机策略作为接入控制策略来使系统性能达到最优,优化目标是最大化系统的收益.基于POMDP模型,我们提出基于观测的策略梯度算法来求解最优策略.最后,使用基于HMM的方法来侦查和预测系统的变化,并且采用动态自适应的方法来更新系统模型机接入控制策略.实验结果显示,和尽力而为的服务策略相比较我们的最优策略获得了较好的性能.
아문위분포식업무계통제출료일충기우모형여예측적접입공제방법.수선,사용부분가관Markov결책과정(POMDP)래위업무계통건모.접착,장복무분배책략융입계통삼수병차사용수궤책략작위접입공제책략래사계통성능체도최우,우화목표시최대화계통적수익.기우POMDP모형,아문제출기우관측적책략제도산법래구해최우책략.최후,사용기우HMM적방법래정사화예측계통적변화,병차채용동태자괄응적방법래경신계통모형궤접입공제책략.실험결과현시,화진력이위적복무책략상비교아문적최우책략획득료교호적성능.
An admission control method was proposed based on modeling and prediction for distributed service systems.First,the partly observable Markov decision process (POMDP) was used to model the service system.Next,the service allocation policy was put into the model parameters and a randomized policy was used as the admission control policy to optimize system performance.The target of optimization is to maximize the system benefits.Based on the POMDP model,an observation-based policy gradient algorithm was proposed to determine the optimal policy.HMM-based method was used to detect and predict the change of system,then the system model and admission policy was updated with the dynamic adaptive method.Experiment results show that compared with the best effort service policy our optimal policy has a better system performance.