Automatic Truss Design with Reinforcement Learning
Authors: Weihua Du, Jinglun Zhao, Chao Yu, Xingcheng Yao, Zimeng Song, Siyang Wu, Ruifeng Luo, Zhiyuan Liu, Xianzhong Zhao, Yi Wu
IJCAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments and ablation studies in popular truss layout design test cases in both 2D and 3D settings. |
| Researcher Affiliation | Collaboration | 1Institute for Interdisciplinary Information Sciences, Tsinghua University 2Tongji University 3East China Architectural Design & Research Institute Co. , Ltd. 4Shanghai Qi Zhi Institute |
| Pseudocode | Yes | Algorithm 1 Auto Truss |
| Open Source Code | No | The paper does not provide any explicit statement about open-sourcing the code for the described methodology, nor does it include a link to a code repository. |
| Open Datasets | Yes | We choose two common 2D test cases in truss layout design [Fenton et al., 2015]: the 10-Bar Cantilever Truss (10Bar) and the 17-Bar Cantilever Truss (17-Bar), as shown in Fig.4. ... We select the Cantilever Sundial Design (Sundial) as the 3D testbed, which follows [Luo et al., 2022a]. |
| Dataset Splits | No | The paper describes its training process (e.g., environment steps for RL training, running multiple seeds) but does not specify a traditional train/validation/test split for a dataset, as the problem involves design generation and refinement on specific test cases rather than model training on a pre-split dataset. |
| Hardware Specification | Yes | More specifically, we run 2e6 iterations in the search stage, which is half the number of iterations in KR-UCT, and 1.5e5 environment steps for RL training, so that the running time of the refinement stage is similar to the search stage with an RTX 3070 GPU. |
| Software Dependencies | No | The paper mentions general techniques and algorithms (e.g., Finite Element Analysis, GNN, SAC algorithm) but does not provide specific software names with version numbers required for reproduction (e.g., Python, PyTorch, specific FEA software versions). |
| Experiment Setup | Yes | More specifically, we run 2e6 iterations in the search stage, which is half the number of iterations in KR-UCT, and 1.5e5 environment steps for RL training, so that the running time of the refinement stage is similar to the search stage with an RTX 3070 GPU. We run 3 seeds for each test case and report the best numbers with the mean numbers and standard deviations. ... The termination criterion of one episode is that the maximum number of 20 actions are performed. We also early terminate an episode if the policy generates 5 invalid layouts within a single episode. |