Advanced deep reinforcement learning for optimizing 3D printing toolpaths: A framework with enhanced agent architectures, Count-Prioritized Replay, and curriculum learning

Ahmed Merze, Fatih Vehbi Çelebi,
Advanced deep reinforcement learning for optimizing 3D printing toolpaths: A framework with enhanced agent architectures, Count-Prioritized Replay, and curriculum learning,
Engineering Science and Technology, an International Journal,
Volume 72,
2025,
102205,
ISSN 2215-0986,
https://doi.org/10.1016/j.jestch.2025.102205.
(https://www.sciencedirect.com/science/article/pii/S2215098625002605)
Abstract: Optimizing toolpaths in 3D printing presents a significant challenge for achieving efficient and high-quality prints. In this study, a novel deep reinforcement learning (DRL) framework is proposed to overcome this problem. A core component of this framework is the proposed agent: Count-Prioritized Replay Deep Q-Network (CPR-DQN). The first key contribution is a developed simulation environment called PrintBoardEnv, which has curriculum learning built in. Secondly, a new method is developed for experience replay called Count-Prioritized Replay (CPR). The third key development is the CPR-DQN agent itself that uses a special architecture which includes features like Implicit Quantile Networks (IQN), Munchausen RL, Dueling, and Noisy Networks. Our agent is trained in two stages respectively, which is an offline pre-training, and then an online training. Our CPR-DQN agent is compared with other agents like DQN, Rainbow DQN, and Beyond The Rainbow (BTR). Furthermore, it is demonstrated that the CPR-DQN agent achieves great performance, highlighting the benefits of the proposed framework for toolpath optimization.
Keywords: Deep reinforcement learning; Toolpath optimization; Count-Prioritized Replay; Curriculum learning; DQN