2 results
Stone masonry design automation via reinforcement learning
- SungKu Kang, Jennifer G. Dy, Michael B. Kane
-
- Article
-
- You have access Access
- Open access
- HTML
- Export citation
-
The use of local natural and recycled feedstock is promising for sustainable construction. However, unlike versatile engineered bricks, natural and recycled feedstock involves design challenges due to their stochastic, sequential, and heterogeneous nature. For example, the practical use of stone masonry is limited, as it still relies on human experts with holistic domain knowledge to determine the sequential organization of natural stones with different sizes/shapes. Reinforcement learning (RL) is expected to address such design challenges, as it allows artificial intelligence (AI) agents to autonomously learn design policy, that is, identifying the best design decision at each time step. As a proof-of-concept RL framework for design automation involving heterogeneous feedstock, a stone masonry design framework is presented. The proposed framework is founded upon a virtual design environment, MasonTris, inspired by the analogy between stone masonry and Tetris. MasonTris provides a Tetris-like virtual environment combined with a finite element analysis (FEA), where AI agents learn effective design policies without human intervention. Also, a new data collection policy, almost-greedy policy, is designed to address the sparsity of feasible designs for faster/stable learning. As computation bottleneck occurs when parallel agents evaluate designs with different complexities, a modification of the RL framework is proposed that FEA is held until training data are retrieved for training. The feasibility and adaptability of the proposed framework are demonstrated by continuously improving stone masonry design policy in simplified design problems. The framework can be generalizable to different natural and recycled feedstock by incorporating more realistic assumptions, opening opportunities in design automation for sustainability.
Automated feedback generation for formal manufacturing rule extraction
- SungKu Kang, Lalit Patil, Arvind Rangarajan, Abha Moitra, Tao Jia, Dean Robinson, Debasish Dutta
-
- Article
- Export citation
-
Manufacturing knowledge is maintained primarily in the unstructured text in industry. To facilitate the reuse of the knowledge, previous efforts have utilized Natural Language Processing (NLP) to classify manufacturing documents or to extract structured knowledge (e.g. ontology) from manufacturing text. On the other hand, extracting more complex knowledge, such as manufacturing rule, has not been feasible in a practical scenario, as standard NLP techniques cannot address the input text that needs validation. Specifically, if the input text contains the information irrelevant to the rule-definition or semantically invalid expression, standard NLP techniques cannot selectively derive precise information for the extraction of the desired formal manufacturing rule. To address the gap, we developed the feedback generation method based on Constraint-based Modeling (CBM) coupled with NLP and domain ontology, designed to support formal manufacturing rule extraction. Specifically, the developed method identifies the necessity of input text validation based on the predefined constraints and provides the relevant feedback to help the user modify the input text, so that the desired rule can be extracted. We proved the feasibility of the method by extending the previously implemented formal rule extraction framework. The effectiveness of the method is demonstrated by enabling the extraction of correct manufacturing rules from all the cases that need input text validation, about 30% of the dataset, after modifying the input text based on the feedback. We expect the feedback generation method will contribute to the adoption of semantics-based technology in the manufacturing field, by facilitating precise knowledge acquisition from manufacturing-related documents in a practical scenario.