Towards Source Code Generation with Recurrent Neural Networks

The evolution of Machine Learning and Data Science disciplines has been rapid duringthe last decade. As computer engineers, we are looking for ways to take advantage of thisevolution.In this diploma thesis we examine the potential of recurrent neural networks to generatesource code, given their effectiveness at handling sequences. We propose two approaches,based on per-character analysis of software repositories. Following appropriate code pre-processing and network training, models generate source code through a stochastic process.We perform static code analysis on model products, in order to examine the performanceof the approaches. We have applied our approach on the JavaScript Language.The analysis shows the great representational power of the recurrent neural networks,but also the inability of our approaches to satisfactorily address the problem of automaticprogramming. Based on these findings, we propose further research directions and waysof exploiting the models that were designed.