Abstract

Over the past few decades, there has been much progress in the field of software development concerning the used technologies as well as the de velopment methodologies. Since the development of traditional algorithms for some problem domains is difficult, the hope for further progress lies in the use of machine learning models.

The integration of machine learning models into software systems introduces new challenges. Though, the development and lifecycle of the models has many similarities with conventional software, the methodologies and technologies do not have reached the same maturity level. The challenges already start with the necessary collaboration of different teams and continue with correct versioning, persisting and serving the models for the production system. Due to debugging is different compared to traditional software, the reproducibility of the original model is of special importance.

Motivated by the problem statements of an existing software system, in this thesis solution approaches for the integration and deployment of machine learning models were worked out. On the one hand best-practices from different companies and projects were researched and on the other hand existing third-party solutions for different parts of the lifecycle relevant for model deployment were analyzed.

A further part of the thesis is about the current architecture of the software system and necessary extensions for the simple integration of the models.