Answers are grounded in this paper and cite their sources.
Reliability-based design optimization (RBDO) needs an explicit objective function, which early-stage design rarely offers, because only a few noisy evaluations exist. Engineers usually do know the direction in which each variable acts; for example, a longer weld cannot lower cost. We encode this knowledge as sign constraints on a Ridge regression surrogate of the objective. Variables with competing physical pathways are split into virtual variables the net effect of which is left to the data. The surrogate replaces only the objective. The reliability constraints stay on the original limit-state functions and are evaluated with the First-Order Reliability Method (FORM), so surrogate error never enters the reliability analysis. On the welded beam and the Golinski speed reducer, the signs help only for variables the effect of which is small relative to the noise, which an effect-to-noise ratio from a first unconstrained fit identifies. On Golinski the paired gain in test R2 reaches 0.01 in 8 of 17 cells, though single datasets can lose accuracy, and with 20 to 100 labels the full-information design is returned in 47% to 86% of the replicates, against 11% to 29%, although mean mass differs by only 0.2% to 0.3%. A wrong sign on a strong variable lowers R2 by 0.42 to 0.83, so doubted signs should be left free.
No takes yet. Share an insight, caveat, or question.
Meselhy et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: