属性驱动的自适应多agent系统设计关键技术研究

Modern software-intensive systems have to operate in open and dynamic environments. The software engineering community is facing grave challenges on the development of self-adaptive systems that are capable of adjusting their behaviors in response to the frequent changes in users’ requirements, the rapid advances in enabling technology, and the vast variety of environmental conditions. It is particularly difficult to design decentralized adaptive systems because of the huge gulf between the macro(system)-level adaptive goal and micro(component)-level adaptive behavior. Agent-oriented software engineering has been regarded as a promising approach to constructing decentralized adaptive systems. This project will develop a property-driven AMAS design method that consists of high-level modeling, simulation and statistical model checking techniques. The proposed project will first study how to formally define macro-level adaptive requirements as self-* properties, and to specify them as the target of the design verification. Second, the project will develop a modeling language that provides language facilities to describe micro-level adaptive strategies. Finally, it will develop a model-based simulation and statistical model checking technique and implement a prototype tool so that the emergent behavior of a designed system can be observed through simulation. Whether the design model satisfies a required self-* property can be quantitatively analyzed through statistical model checking. The proposed method, therefore, will bridge the gap between the micro-level design of adaptive strategies and macro-level goal of decentralized adaptive systems.

由于计算系统的运行环境日益开放，软件工程领域面临的重要挑战之一是如何开发能改变自身结构和行为、适应运行条件变化和用户需求演化的自适应软件。大规模开放式应用适合采用分散式自适应控制，基于agent的方法是开发分散式自适应系统的一种有效途径。如何通过设计个体agent的自适应行为实现系统的自适应目标，仍是有待解决的问题。本项目将自适应属性作为贯穿设计过程的驱动因素，提出一种融合建模、模拟和统计模型检验的自适应多agent系统设计方法。提供涵盖多种self-*属性的定义框架，支持将自适应目标形式地规约为self-*属性。建模语言具有完整的形式化语义，提供适用于描述微观层面自适应策略的语言设施。设计模型的动态模拟有助于展现由agent交互产生的系统级涌现行为，统计模型检验能够定量地分析设计模型是否满足self-*属性。因此，该方法能够弥补微观层面的自适应策略设计与宏观层面的自适应目标之间的鸿沟。