From Code to Knowledge to Software

点击翻译按钮获取中文摘要

In programming languages, we have some well understood technologies: interpretation, compilation, partial evaluation, code generation, etc. In fact, for most Domain Specific Languages, these can be automatically generated. If we furthermore understand software as forming families, commonalities and variabilities can be encoded in a generator. When properly done, the generator not only highlights possible design choices, it also provides for traceability between source knowledge and eventual software artifacts -- such as cross-referencing each computed quantity with its derivation in the design/requirements document(s). Traceability is onerous via traditional methods, and prohibitive to do after the fact. Regulated industries, such as nuclear, medical, aerospace, require traceability as part of the certification process. Current development methods for software which can be certified are document driven, and involve a large amount of knowledge duplication. This duplication is both expensive and hinders traceability. Generative methods have been very successfully applied to code in certain domains like scientific computation, user interfaces, operating systems and graphics. However, code is only a small part of all artifacts which makes up software. Thus we propose to generalize the generative, family-based approach to encompass all artifacts. By design, these share a lot of knowledge: requirements specifications, design documents, code, user manuals, tests, etc, all should say the same thing. The main long term goal of the research is to foster substantial long-term productivity increases by using generative techniques to avoid information duplication. To achieve this, we will develop a sequence of domain specific languages 1) of (requirements) documents, 2) of "generic" object-oriented languages, 3) of design and implementation choices, and 4) of algorithmic knowledge. This is not a silver bullet: not all problem domains are sufficiently well understood to be captured in this way. The research will focus on domains where a large body of knowledge exists. More generally, the problem domain and its software solution must be based on well-understood theory, where the translation from requirements to design to code, but also tests, user manuals, etc, must all be well understood, i.e. that the science and engineering of such software is established. These domains generally fall under the umbrella of scientific computation. The proposed research will form the foundations of long-term productivity growth in software development, especially in scientific computation, and train 20 HQP in these methods. Our methodology emphasizes working on specific examples of engineering software as a means of grounding our research. Productivity growth is critical to stay ahead of the coming "software crisis" (as the recent article in The Atlantic calls it).