GISMO: Genetic Improvement of Software for Multiple Objectives

人类发现很难开发出平衡许多相互竞争和冲突的运营目标的系统。即使实现一个目标，也需要自动支持。例如，对优化编译代码大小和速度的技术的研究已经有了漫长而丰富的历史。不幸的是，速度和规模只是下一代软件系统必须达到的众多目标中的两个。新兴的计算应用范例不仅要求系统可靠、紧凑和快速，而且要求系统优化许多不同的竞争和冲突目标，如响应时间、吞吐量和资源消耗。不能指望人类以最佳方式平衡这些多重相互竞争的约束，并可能错过潜在的有价值的解决方案。因此，需要能够自动创建代码以平衡许多相互冲突的目标的技术，或者能够为寻求这样做的人提供支持的技术。Gismo项目试图做到这两点。它将开发自动化技术，以生产符合新定义目标的现有系统组件的新版本。在并行运行旧组件和新组件一段时间后，程序员可以决定采用新发展的组件。然而，Gismo方法的美丽之处在于，它并不坚持程序员必须接受进化的解决方案才能有用。程序员还可以使用Gismo来探索多目标候选解决方案空间，深入了解通过平衡几个相互竞争的约束可以实现什么。

Humans find it hard to develop systems that balance many competing and conflicting operational objectives. Even meeting a single objective requires automated support. For example, there has been a long and rich history of research into techniques to optimise compiled code size and speed. Unfortunately, speed and size are but two of many objectives that the next generation of software systems will have to meet. Emergent computing application paradigms require systems that are not only reliable, compact and fast, but also which optimise many different competing and conflicting objectives such as response time, throughput and consumption of resources. Humans cannot be expected to optimally balance these multiple competing constraints and may miss potentially valuable solutions. Techniques are therefore required that can either automatically create code that balances many conflicting objectives or that can provide support to the human who seeks to do so. The GISMO project seeks to do both. It will develop automated techniques to produce new versions of components of existing systems that meet newly defined objectives. After a period of running the old and new component in parallel, the programmer may decide to adopt the newly evolved component. However, the beauty of the GISMO approach is that it does not insist that the programmer must accept the evolved solution in order to be useful. The programmer can also use GISMO to explore the multi-objective candidate solution space, gaining insight into what can be achieved by balancing several competing constraints.