Automating Code Generation for MDE using Machine Learning

The overall aim of our research is to improve the techniques for synthesizing code generators in the Model-Driven Engineering (MDE) context. Code generation is one of the main elements of Model-Driven Engineering, involving transformation from specification models to produce executable code. A code generator is designed to reduce the manual program construction work used to implement a software system, but building a code generator itself still currently needs much manual effort. Meanwhile, existing code generators are typically not flexible to adjust for changing development requirements and are hard to reuse for different target languages. Therefore, we aim to provide techniques to improve the process of building code generators, and let them be more reusable. Currently, we researched the related new and traditional approaches for generating code and projects using AI for program translation, code completion or program generation. Based on this research we decided to focus on a symbolic machine learning method related to the programming-by-example concept to build code generators. We use this “Code Generation By Example” (CGBE) concept with tree-to-tree structure mappings as the information format. CGBE has good performance in terms of training dataset size and time when applied to learning a UML-to-Java code generator, but further work is needed to extend it to generate different programming languages and to evaluate these cases, and to handle the optimisation of generated code.