Automated program improvement with reinforcement learning and graph neural networks

Nataša Sukur

doi:10.1007/s00500-023-08559-1

2023-06-03

PRE

DOI：10.1007/s00500-023-08559-1

原文链接

原文求助

摘要

En 中文

Automated software transformations and the process of automated program repair and improvement are an important aspect of modern software engineering practices. In this paper, we describe a system which uses a graph-based deep learning model that can be trained to automatically transform and improve computer programs. By operating on language-agnostic, universal graph-like structures easily extractable from source code files (abstract syntax trees), the deep learning agent learns which transformations should be effectively applied to various structures recognized in the source code in order to improve it. By defining a metric which we want to improve and introducing an optimization task-a reinforcement learning setting, an agent learns to automatically apply a chain of transformations to the program, drastically improving it. While similar program improvement processes exist, they exclusively use exhaustive search algorithms to try all the possible code transformations which is a long process susceptible to local optimum issues. Our solution aims to model and embed structural knowledge about the programs being transformed which greatly helps the agent to choose best possible code transformations to apply. Elements of the approach we present in this paper are further applicable not just to automatic software improvement tasks, but also to other code-related tasks.

Keyword:

Automatic program transformation

Software evolution

Deep reinforcement learning

Graph embeddings

Graph reinforcement learning