From 7bb88cff6c1d6a04c561a7b78743eb7752f66024 Mon Sep 17 00:00:00 2001 From: =?utf8?q?J=C3=A9r=C3=B4me=20Benoit?= Date: Thu, 17 Sep 2026 21:31:37 +0200 Subject: [PATCH] docs(quickadapter): document selection normalization bounds and schema --- README.md | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/README.md b/README.md index a0398a5..afb2979 100644 --- a/README.md +++ b/README.md @@ -194,6 +194,16 @@ Operational constraints: Singular covariance eigenvalues are floored at the largest eigenvalue times the square root of machine epsilon. These two metrics do not support custom objective weights; configured weights warn and fall back to uniform weights. +- Objective weights are validated once and normalized by maximum then sum; + scale-invariant selection survives any positive rescaling of the same + weights. An all-zero configuration falls back to uniform weights. +- Combined-metric softmax aggregation is numerically stable for any finite + positive temperature and coefficients; log coefficients are applied after + temperature division. Boolean metric coefficients are ignored. +- Invalid label metric values of any JSON type follow the documented + warning/fallback policy instead of raising. +- Reversed label HPO bounds publish their validated fallback to the shared + configuration, so the label objective receives consistent ranges. - With `space_reduction=true`, `space_fraction=0` fixes numerical search ranges at valid previous best values; it does not freeze every categorical choice. Saved HPO suggestions reconstruct derived estimator parameters for final fits: @@ -203,6 +213,10 @@ Operational constraints: signed labels. Explicit `"lognormal"` requires finite, strictly positive transformed training and validation labels; incompatible HPO trials are pruned, while direct fits raise an error. +- Changing objective-weight normalization, softmax stabilization, metric-type + validation, or bound publication resets label HPO selection state + (selection schema 4). The persisted best-params JSON layout remains + independently versioned. ### Backtest evaluation protocol -- 2.53.0