Table 1 Summary table for the literature review.
From: Reinforcement learning-driven dynamic optimization strategy for parametric design of 3D models
Author(s) | Approach/Model | Pros | Cons |
|---|---|---|---|
Brown et al.16 | Deep RL-based topology optimization for 2D designs | Achieves comparable or better performance than gradient-based methods; automates element removal | Scalability issues in 3D; high computational cost |
Tskhondia17 | GA + RL hybrid topology optimization using PPO and HRL | Improved efficiency; reduced elapsed optimization time; better than standalone GA or RL | Poor scalability for high-resolution grids; possible inefficiencies in algorithm transfer |
Kolodiazhnyi et al.18 | Multi-modal CAD reconstruction using GRPO and vision-language models | Strong performance across benchmarks; combines multiple data modalities | Requires large synthetic datasets; high training complexity |
Rochefort-Beaudoin et al.19 | SOgym: RL-based structural optimization with physical constraints | Supports model-free and model-based agents; mesh-independent feature mapping; good conformity to constraints | Slower learning rates; long training periods required |
Wang & Hu20 | RL for airfoil shape optimization with PDE solvers | High accuracy through physics-based modeling; responsive to geometric changes | Very computationally intensive; complex solver–RL integration |
He & Ciocarlie21 | MORPH: RL with differentiable hardware co-optimization | Enables long-horizon tasks; flexible proxy-based simulation | Proxy may not accurately represent real hardware; limited real-world validation |
Tang & Hong22 | RL-based task migration for thermal regulation in 3D architectures | Reduces peak temperature by 31%; efficient thermal management | Increased complexity for large core systems; scalability concerns |
RM Sakiyama et al.23 | Parametric MOO for ventilation and energy efficiency with sensitivity analysis | Significant improvements in NVE and THL; reduced design variables | Dependent on simulation accuracy; climate-specific applicability |
Behzadi & Ilies24 | Deep transfer learning for real-time 3D topology optimization | Enables real-time inference; avoids iterative solvers; suitable for parametric design exploration | Requires large precomputed datasets; limited generalization; lacks explicit constraint handling and hierarchical decision structure |
