Streamlining AWS DeepRacer Model Training: A Python Library for Jupyter Notebook Integration with DeepRacer For Cloud

Abstract

The open-source DeepRacer for Cloud (DRfC) framework democratizes access to Reinforcement Learning (RL) experimentation, but its command-line interface imposes a high usability barrier, especially for new users. This thesis aimed to design, implement, and empirically validate a highlevel abstraction layer for DRfC to simplify its workflow and improve the user experience. The solution, a Python library named drfc_manager, was developed to orchestrate DRfC operations within an interactive Jupyter Notebook environment. Its effectiveness was evaluated through a comparative usability study with 21 participants, who performed a set of standardized tasks using both the proposed library and the traditional script-based method. The results demonstrate that the drfc_manager library significantly reduces task completion time; for instance, ’Training Initiation’ was completed in under 2 minutes by 85.7% of library users, in contrast to 90.5% of baseline users requiring 5 minutes or more. The solution was perceived as vastly easier to use, culminating in an overall user preference of 95.2% for the library. This work contributes an empirically validated, publicly available open-source tool that reduces the complexity of DRfC, making RL experimentation more accessible and efficient. The findings reinforce the importance of user-centric interfaces for the adoption of complex software platforms.