Automated Discovery of Emergent Misbehaviour

计算模型本质上是旨在模拟自然系统的计算机程序，例如蚁群，皮肤组织的形成或全球经济。科学家和工业家使用计算模型来帮助他们理解所模拟的自然系统，进行预测，并预测某些变化对系统的影响。计算机模型应用的一个典型例子可能是预测国有化北方岩银行对英国经济的影响。基于代理的模型是从系统中的主要参与者的角度开发的计算模型，例如，公民起义模型中的恐怖分子，或选举中的选民。从模拟模型中得到的观察结果可以用所涉及的个人来理解和解释，回答诸如“为什么这个国家的这个地区会有内乱？'，'为什么这个政党在这里很受欢迎？或者“为什么这里会长肿瘤？'.从由一组数学方程组成的“自上而下”的建模方法中，可以说很难做出这样的推论。由于预测和科学发现都依赖于正确实施的模型，因此没有正确测试模型的后果可能非常严重，过去已经使公司损失了数百万英镑。然而，传统的软件测试策略，这需要一个“分而治之”的方法，很难适用于基于代理的模型。在模拟中，代理的交互通常是随机的，会产生复杂的模式和行为。因此，很难预测哪些原因会直接导致哪些影响。这里提出的研究旨在使用智能搜索技术测试基于代理的模型，具体目的是“归巢”在以前的模拟运行中没有暴露的模型行为。通过这种方式，搜索过程将鼓励在不太可能或考虑不周的情况下测试模型，其中模型的行为可能偏离预期。为了做到这一点，搜索过程将通过扩展一种称为不变检测的技术，建立一个模型在模拟中所做的抽象图片。不变量是总是被发现为真的陈述。该搜索的主要目的是伪造从过去的模拟中生成的不变量，以证明模型的新行为。这些可能是罕见的或意外的行为，可能以前没有测试过，因此可能是软件错误潜伏的情况。

Computational models are essentially computer programs intended to simulate a natural system, such as an ant colony, the formation of skin tissue or the global economy. Scientists and industrialists use computational models to help develop their understanding of the natural system being modelled, to make forecasts, and to predict the impact of some change to the system. A topical example of an application of a computer model might be to predict the impact of nationalising the Northern Rock bank on the UK economy.Agent-based models are computational models which have been developed from the point of view of the main actors in the system, e.g. a terrorist in a model of civil uprising, or a voter in an election. Observations made from simulating the model can be understood and explained in terms of the individuals involved, answering questions such as 'why is there civil disturbance in this area of the country?', 'why is this political party popular here?' or 'why has a tumour developed here?'. Such inferences are arguably difficult to make from 'top-down' approaches to modelling, which are composed of a set of mathematical equations.As predictions and scientific discoveries are reliant on the models being implemented correctly, the consequences of not properly testing a model can be extremely serious, and have cost companies several millions of pounds in the past. However, traditional software testing strategies, which take a 'divide and conquer' approach, are difficult to apply to agent-based models. The interaction of agents in a simulation, often at random, produces complex patterns and behaviours. Thus, it is difficult to predict which causes will lead directly to which effects. The proposed research here intends to test agent-based models using intelligent search techniques, with the specific intention of 'homing in' on behaviours of the model that have not been exposed in previous simulation runs. In this way, the search process will encourage testing of the model in unlikely or ill-conceived situations, where the model's behaviour may diverge from that intended. In order to do this, the search process will build up an abstract picture of what the model is doing in simulation through extension of a technique known as invariant detection. Invariants are statements that are always found to be true. The search will essentially aim to falsify generated invariants generated from past simulations in order to demonstrate new behaviours of the model. These are likely to be rare or unexpected behaviours that may not have been previously tested, and thus possible instances where software errors may be lurking.