A dynamic domain decomposition method for solving coupled multiphysics problems based on a distributed software agent environment

该项目的目标是在运行时优化由软件代理计算的复杂三维耦合场问题的高效和灵活的计算过程。以前的工作表明，在多代理系统中使用多个软件代理，使用专门的数值场计算算法，如有限元法或边界元法，提供了有效和灵活的计算。因此，模拟是静态划分的问题类，如所包括的物理，其域或可用资源。分离的顺序基于部分问题之间的预期耦合的强度。这代表了在分布式资源上进行计算的常用方法。考虑到可用的计算资源、计算代理的能力和它们的工作量，在软件代理系统内分散地进行部分问题的计算以及考虑相关的耦合。在运行时期间，基于信息，例如收敛行为或部分计算的当前状态，可以验证分裂以及耦合层次。由于部分问题之间的非线性耦合，并且由于每个分布式集群的不确定性行为，动态耦合层次结构在数值上有利于迭代分布式计算方法的收敛。基于现有的代理系统的耦合的动态考虑是可能的，通过基于智能和学习决策的主动协调。因此，学习系统的实现使用用于耦合系统的可存储参数的知识库。这也包括分布式、分散式架构，在这种架构中，知识不是集中产生的，而是在不同的单位中临时产生的。考虑到从知识库的建议是通过调整计算过程中的数值迭代求解过程。对所取得的成果进行最后评估，完成了该项目。它是做了特别注意实现优化的耦合系统计算与分散和自主代理。另一点是学习系统所实现的可比性和增益。

The optimization of the efficient and flexible calculation process of complex three-dimensional coupled field problems, calculated by software agents, during run time is the goal of this project. Previous work shows, that the use of multiple software agents within a multi agent system using specialized algorithms of numerical field calculation such as the finite element method or boundary element method provide the efficient and flexible calculation. Therefore, the simulation is statically split on classes of problems like the included physics, its domains or available resources. The order of the separation is based on the strength of the expected coupling between the partial problems. This represents the usual approach to computations on distributed resources. The calculation of the partial problems as well as considering related couplings are done decentralized within the software agent system considering available computing resources, the capabilities of the calculation agent and their workload. A validation of the splitting - and thus also the coupling hierarchy is possible during runtime based on information, such as the convergence behaviour or the current state of partial calculations. Due to non-linear coupling between the partial problems, and due to the nondeterministic behaviour of every distributed cluster, a dynamic coupling hierarchy is numerically beneficial for the convergence of an iterative distributed calculation approach. Based on the existing agent system a dynamic consideration of the coupling is possible by an active coordination based on intelligent and learned decisions. Accordingly, the implementation of a learning system using a knowledge base for storable parameters of a coupled systems. This also includes the distributed, decentralized architecture where knowledge is not generated centrally, but on different units available only temporarily. Taken recommendations from the knowledge base into account is done within the numerical iterative solution procedure by adapting the calculation process. The final evaluation of the achieved results completes this project. It is done with special attention to the achieved optimization of coupled systems calculation calculated with decentralized and autonomously acting agents. An additional point is the comparability and gain achieved by the learning system.