Reinforcement learning-driven dynamic optimization strategy for parametric design of 3D models

The concept of parametric design is changing the way 3D modeling works, allowing precise manipulation of complex forms in the areas of architecture, digital fabrication, and product design. However, exploring and optimizing large coupled spaces of parameters remains a significant computational challenge. We present a new, Hierarchical Reinforcement Learning based Dynamic Optimization Strategy (HRL-DOS), which decomposes the parametrized design process into a series of multi-level subproblems. The high-level policy determines the global direction of the design while the low-level policy adapts individual parameters, responding to changes from multiple performance criteria (structural stability, geometric efficiency, and fabrication constraints). The hierarchical approach provides greater efficiency in learning and computational scaling in a complex design environment. Experimental tests on benchmark 3D modeling tasks revealed a 27% improvement in convergence and 18% improvements in quality of the model, relative to simple heuristic or gradient-based optimizations. In addition, HRL-DO permits adaptability in real-time, and the approach can potentially translate to various domains, including automated form-finding for architectural structures, generative design of products, or intelligent computer-aided design (CAD) systems. Through the use of HRL, we have developed a new and adaptive approach for the additional automation of parametric design tasks in the future.