Browse Teach Curate

Gumbel-Softmax Gradient-Based Optimization of Categorical Design Variables in Structural Design

A gradient-based structural-optimization method that uses the Gumbel-Softmax reparameterization to draw differentiable samples from categorical distributions, making discrete, unordered design choices (e.g. cross-section profiles, materials) amenable to sensitivity analysis so they can be optimized jointly with continuous variables at large scale, avoiding the exponential cost of gradient-free search.

B9 Text

2501.00258

This paper proposes a gradient-based optimizer for the optimal design of large-scale frame structures with mixed categorical (discrete, unordered choices such as cross-sectional profile or material) …