Reorganization of Agent Networks with Reinforcement Learning based on Communication Delay
Reorganization of Agent Networks with Reinforcement Learning based on Communication Delay
复制标题
基于通信延迟的强化学习智能体网络重组
DOI：
10.1109/wi-iat.2012.105
复制
发表时间：
2012
期刊：
影响因子：
--
通讯作者：
Kazuki Urakawa and Toshiharu Sugawara
中科院分区：
文献类型：
--
作者：
桂木敦史;櫻井祐子;岩崎敦;横尾真;Kazuki Urakawa and Toshiharu Sugawara
提出了基于通信延迟的强化学习和Agent网络重组的Agent网络任务分配的团队形成方法。在分布式环境中，如Internet应用程序（如网格计算和面向服务的计算）中的任务通常是通过执行多个子任务来实现的。这些子任务是以自底向上的方式按需构建的，并且必须使用具有每个子任务所需的能力和计算资源的适当代理来完成。因此，在这种系统中，高效和有效地将任务分配给合适的代理是一个关键问题。在我们的模型中，这个分配问题制定为面向任务的域中的代理的团队形成。从这个角度出发，进行了一些研究，其中包括学习和重组。本文的目的是从两个角度扩展传统的方法。首先，我们提出的方法只使用本地可用的信息进行学习，从而使这种方法适用于真实的系统。其次，我们引入了消除链接以及在代理网络中的链接，以提高学习效率。实验表明，与传统方法相比，这种扩展可以大大提高团队组建的效率。我们还表明，它可以使代理网络适应环境的变化。
We propose the team formation method for task allocations in agent networks by reinforcement learning based on communication delay and by reorganization of agent networks. A task in a distributed environment like an Internet application, such as grid computing and service-oriented computing, is usually achieved by doing a number of subtasks. These subtasks are constructed on demand in a bottom-up manner and must be done with appropriate agents that have capabilities and computational resources required in each subtask. Therefore, the efficient and effective allocation of tasks to appropriate agents is a key issue in this kind of system. In our model, this allocation problem is formulated as the team formation of agents in the task-oriented domain. From this perspective, a number of studies were conducted in which learning and reorganization were incorporated. The aim of this paper is to extend the conventional method from two viewpoints. First, our proposed method uses only information available locally for learning, so as to make this method applicable to real systems. Second, we introduce the elimination of links as well as the generation of links in the agent network to improve learning efficiency. We experimentally show that this extension can considerably improve the efficiency of team formation compared with the conventional method. We also show that it can make the agent network adaptive to environmental changes.