移动边缘计算(Mobile Edge Computing,MEC)将计算与存储资源部署到网络边缘,用户可将移动设备上的任务卸载到附近的边缘服务器,得到一种低延迟、高可靠的服务体验.然而,由于动态的系统状态和多变的用户需求,MEC环境下的计算卸载与资源...移动边缘计算(Mobile Edge Computing,MEC)将计算与存储资源部署到网络边缘,用户可将移动设备上的任务卸载到附近的边缘服务器,得到一种低延迟、高可靠的服务体验.然而,由于动态的系统状态和多变的用户需求,MEC环境下的计算卸载与资源分配面临着巨大的挑战.现有解决方案通常依赖于系统先验知识,无法适应多约束条件下动态的MEC环境,导致了过度的时延与能耗.为解决上述重要挑战,本文提出了一种新型的基于深度强化学习的计算卸载与资源分配联合优化方法(Joint computation Offloading and resource Allocation with deep Reinforcement Learning,JOA-RL).针对多用户时序任务,JOA-RL方法能够根据计算资源与网络状况,生成合适的计算卸载与资源分配方案,提高执行任务成功率并降低执行任务的时延与能耗.同时,JOA-RL方法融入了任务优先级预处理机制,能够根据任务数据量与移动设备性能为任务分配优先级.大量仿真实验验证了JOA-RL方法的可行性和有效性.与其他基准方法相比,JOA-RL方法在任务最大容忍时延与设备电量约束下能够在时延与能耗之间取得更好的平衡,且展现出了更高的任务执行成功率.展开更多
针对近端策略优化(PPO)算法难以严格约束新旧策略的差异和探索与利用效率较低这2个问题,提出一种基于裁剪优化和策略指导的PPO(COAPG-PPO)算法。首先,通过分析PPO的裁剪机制,设计基于Wasserstein距离的信任域裁剪方案,加强对新旧策略差...针对近端策略优化(PPO)算法难以严格约束新旧策略的差异和探索与利用效率较低这2个问题,提出一种基于裁剪优化和策略指导的PPO(COAPG-PPO)算法。首先,通过分析PPO的裁剪机制,设计基于Wasserstein距离的信任域裁剪方案,加强对新旧策略差异的约束;其次,在策略更新过程中,融入模拟退火和贪心算法的思想,提升算法的探索效率和学习速度。为了验证所提算法的有效性,使用MuJoCo测试基准对COAPG-PPO与CO-PPO(PPO based on Clipping Optimization)、PPO-CMA(PPO with Covariance Matrix Adaptation)、TR-PPO-RB(Trust Region-based PPO with RollBack)和PPO算法进行对比实验。实验结果表明,COAPG-PPO算法在大多数环境中具有更严格的约束能力、更高的探索和利用效率,以及更高的奖励值。展开更多
To facilitate users to access the desired information, many researches have dedicated to the Deep Web (i.e. Web databases) integration. We focus on query translation which is an important part of the Deep Web integr...To facilitate users to access the desired information, many researches have dedicated to the Deep Web (i.e. Web databases) integration. We focus on query translation which is an important part of the Deep Web integration. Our aim is to construct automatically a set of constraints mapping rules so that the system can translate the query from the integrated interface to the Web database interfaces based on them. We construct a concept hierarchy for the attributes of the query interfaces, especially, store the synonyms and the types (e.g. Number, Text, etc.) for every concept At the same time, we construct the data hierarchies for some concepts if necessary. Then we present an algorithm to generate the constraint mapping rules based on these hierarchies. The approach is suitable for the scalability of such application and can be extended easily from one domain to another for its domain independent feature. The results of experiment show its effectiveness and efficiency.展开更多
文摘移动边缘计算(Mobile Edge Computing,MEC)将计算与存储资源部署到网络边缘,用户可将移动设备上的任务卸载到附近的边缘服务器,得到一种低延迟、高可靠的服务体验.然而,由于动态的系统状态和多变的用户需求,MEC环境下的计算卸载与资源分配面临着巨大的挑战.现有解决方案通常依赖于系统先验知识,无法适应多约束条件下动态的MEC环境,导致了过度的时延与能耗.为解决上述重要挑战,本文提出了一种新型的基于深度强化学习的计算卸载与资源分配联合优化方法(Joint computation Offloading and resource Allocation with deep Reinforcement Learning,JOA-RL).针对多用户时序任务,JOA-RL方法能够根据计算资源与网络状况,生成合适的计算卸载与资源分配方案,提高执行任务成功率并降低执行任务的时延与能耗.同时,JOA-RL方法融入了任务优先级预处理机制,能够根据任务数据量与移动设备性能为任务分配优先级.大量仿真实验验证了JOA-RL方法的可行性和有效性.与其他基准方法相比,JOA-RL方法在任务最大容忍时延与设备电量约束下能够在时延与能耗之间取得更好的平衡,且展现出了更高的任务执行成功率.
文摘针对近端策略优化(PPO)算法难以严格约束新旧策略的差异和探索与利用效率较低这2个问题,提出一种基于裁剪优化和策略指导的PPO(COAPG-PPO)算法。首先,通过分析PPO的裁剪机制,设计基于Wasserstein距离的信任域裁剪方案,加强对新旧策略差异的约束;其次,在策略更新过程中,融入模拟退火和贪心算法的思想,提升算法的探索效率和学习速度。为了验证所提算法的有效性,使用MuJoCo测试基准对COAPG-PPO与CO-PPO(PPO based on Clipping Optimization)、PPO-CMA(PPO with Covariance Matrix Adaptation)、TR-PPO-RB(Trust Region-based PPO with RollBack)和PPO算法进行对比实验。实验结果表明,COAPG-PPO算法在大多数环境中具有更严格的约束能力、更高的探索和利用效率,以及更高的奖励值。
基金Supported by the National Natural Science Foundation of China (60573091)the Natural Science Foundation of Beijing(4073035)the Key Project of Ministry of Education of China (03044)
文摘To facilitate users to access the desired information, many researches have dedicated to the Deep Web (i.e. Web databases) integration. We focus on query translation which is an important part of the Deep Web integration. Our aim is to construct automatically a set of constraints mapping rules so that the system can translate the query from the integrated interface to the Web database interfaces based on them. We construct a concept hierarchy for the attributes of the query interfaces, especially, store the synonyms and the types (e.g. Number, Text, etc.) for every concept At the same time, we construct the data hierarchies for some concepts if necessary. Then we present an algorithm to generate the constraint mapping rules based on these hierarchies. The approach is suitable for the scalability of such application and can be extended easily from one domain to another for its domain independent feature. The results of experiment show its effectiveness and efficiency.