Automatic model repair using reinforcement learning

Abstract

When performing modeling activities, the chances of breaking a model increase together with the size of development teams and number of changes in software specifications. One option to prevent and repair broken models is to automatize this process with a software tool, using techniques like Machine Learning (ML). Despite its potential to help fixing models, it is still challenging to apply ML techniques in the modeling domain due to the lack of available datasets. However, some ML branches could offer promising results since they do not require initial training data, such as Unsupervised and Reinforcement Learning (RL). In this paper we present a prototype tool for automatic model repairing by using RL algorithms. These algorithms could potentially reach model repairing with human-quality without requiring supervision. To conduct an initial evaluation we have tested our prototype with a set of broken models and studied its execution and results.