Repository navigation
Conversation
…entiable surrogates The exact nbc.* group now has Torch calculators. The graph is selected under no_grad and the selected edge lengths are recomputed from x, so the four distance features carry a design gradient. With a fixed design they stay locally constant in y, because the fitness ordering picks the edges and the spacing does not; nbc.nb_fitness.cor is the exception and differentiates through y at a locally fixed graph. Add nbc_soft.*, five differentiable surrogates, in both backends. A smooth first-success distribution over distance-ordered candidates replaces the discrete "first strictly better neighbour" search, which restores a usable gradient with respect to y for all five statistics, including the distance ratios that the exact group cannot provide. Weights and the no-better mass partition each row exactly, and every soft quantity converges to its exact counterpart as the temperature approaches zero for distinct objective values. The gates rank the objectives rather than min-max scaling them, and this is fixed rather than configurable. Measured on BBOB d=2 with n=250d across six functions, min-max gates diverged from the exact features 2.4 to 7.8 times further at every temperature tested, and both are invariant to a positive affine change of y, so nothing was traded away. The default temperature of 0.003 is the measured balance point: divergence of nbc_soft.nn_nb.sd_ratio sits near one standard deviation of the exact feature's own design-to-design sampling noise, while a tenth of it concentrates 98% of the gradient norm onto 5% of the observations against 89% at the default. Two documented gaps: equal objective values gate at 1/2 at every temperature, so plateaus do not reproduce the exact equal-fitness fallback, and a surrogate value is only comparable to a target computed at the same temperature and sample size. Report the exact group on fresh designs. Information content deliberately keeps no Torch calculator, so it stays usable as a held-out family that gradient training cannot reach.
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.