Testing multi-agent systems for unwanted behavior

自适应或与其他系统交互的计算机系统对软件测试方法提出了新的挑战。虽然这样的系统提供了更多自治的可能性(以及对人类管理的较少需求)，以及提供单一系统不可能提供的能力的紧急行为的可能性，但这些同样的可能性也可能导致负面和意想不到的影响。一个系统可能以错误的方式适应，或者当几个系统试图彼此交互时，我们可能会出现错误行为。这项研究是关于新的测试方法，能够发现一个系统或一组系统(多智能体系统)的这种不想要的行为。这将有助于确保我们能够利用多代理系统的优势，而不会出现特殊的安全问题。 与测试系统在给定场景中是否显示正确响应的“传统”测试方法不同，我们的测试方法能够发现系统显示不需要的行为的场景。使用我们的方法，我们已经成功地在FIFA-99足球比赛中发现了不想要的行为(即，用户操作总是导致对计算机玩家得分)，在多代理系统班级中由学生编写的一组救援代理中(即，创建代理以使学生的代理冻结的方式与学生代理交互的行为)，以及在港口巡逻和拦截政策中(即恐怖船只的行为，导致它们在防御者正确遵守策略的情况下仍达到给定目标)。通过建议的研究计划，我们调查了对我们的基本方法的扩展，这些扩展扩展了我们的基本流程，允许引入子目标，允许集成关于测试、应用领域和被测试系统的各种知识，并创造了使先前的测试运行影响下一次测试运行的可能性。这项研究将使我们能够开发一个扩展的通用框架，用于测试针对多代理特定问题的多代理系统，以及如何为具体的多代理系统实例化该框架的指导方针。

Computer systems that adapt themselves or interact with other systems provide new challenges to software testing methods. While such systems offer the possibility for more autonomy (and less need for human administration) and the possibility for emergent behaviors that provide abilities that are not possible for a single system, these same possibilities can also result in negative and unexpected effects. A system might adapt in the wrong way, or we might have emergent misbehavior when several systems try to interact with each other. This research proposal is about new testing methods that are able to find such unwanted behavior of a system or a group of systems (a multi-agent system). This will help to ensure that we can make use of the advantages of multi-agent systems without special safety concerns. In contrast to "conventional" testing methods that test if a system shows the right response in given scenarios, our testing method is able to find scenarios in which a system is showing an unwanted behavior. Using our approach, we were already successful in finding unwanted behavior in the FIFA-99 soccer game (namely user actions that always resulted in scoring a goal against the computer player), in a group of rescue agents written by students in a multi-agent system class (namely creating behaviors of agents interacting with the students' agents in such a manner that the students' agents froze), and in harbour patrol and interception policies (namely behaviors for terrorist boats that resulted in them achieving a given objective despite the defenders correctly following the policies). With the proposed research program, we investigate extensions to our basic method that expand our basic process to allow for the introductions of subgoals, that allow the integration of various knowledge about testing, the application area and the tested system, and that create the possibility to have previous test runs influence the next test run. This research will enable us to develop an extended general framework for testing multi-agent systems for multi-agent specific problems together with guidelines on how to instantiate this framework for concrete multi-agent systems.