Dataset and Reproduction Package for A Guarded Trainable Exponent Operator: Structure-Matched Extrapolation, Loss Targets, and Numerical Boundaries
Description
This repository contains the complete experimental datasets, fitted model checkpoints, predictions, metric tables, and reproduction scripts supporting the benchmark audit of a guarded trainable-exponent operator for positive-domain power-law regression. Research Hypothesis & Objectives: We evaluate whether a compact guarded trainable-exponent product, f(x) = exp(clip(b + sum(q_j * log(max(x_j, epsilon))), -L, L)), can effectively capture positive power-law relationships, and systematically investigate how structural constraints (individual vs. shared exponents), loss targets (log-loss vs. raw-MSE), and numerical safeguards (epsilon-floor, L-cap) govern predictive accuracy, out-of-domain (OOD) extrapolation, and numerical stability. Dataset Contents & Experimental Structure: 1. Synthetic Power-Law Grid (10,368 Primary Fits): Evaluates four power families (product, ratio, unequal powers, shared power) across sample sizes n in {32, 64, 256}, log-noise SD sigma in {0, 0.01, 0.05, 0.1, 0.2, 0.5}, and 24 paired seed clusters across six models (individual q, shared q, fixed q, factorized q, log OLS, RealNPU, MLP16). 2. Diagnostic & Non-Power Grid (3,096 Fits): Evaluates model behavior under misspecification on positive additive, mixed, and smooth non-power targets, alongside 864 OLS guard sweeps (epsilon in {1e-12..1e-4}, L in {2..50}, FP16/FP32/FP64) and 1,152 mixed-component target fits. 3. Real-World Benchmarks: Grouped 22-hull out-of-fold evaluations on Yacht Hydrodynamics (308 observations, 330 model states) and 22-distance-fold evaluations on NIST Chwirut2 (54 observations, official start sensitivities). 4. Controlled Baselines (2,520 Adam Fits & GINN-LP/MSN Comparisons): Objective-matched evaluations comparing guarded TE models against wider MLPs, RealNPU, Müntz–Szász Networks (MSN), and frozen GINN-LP author code under noisy-validation checkpoint selection. Key Findings & Data Interpretation: - Constraint Alignment: Freeing individual exponents is essential for unequal powers, whereas enforcing the shared-exponent constraint on shared-power targets reduces mean OOD Weighted Absolute Percentage Error (WAPE) from 0.0920 to 0.0605. - Safeguard Sensitivity: Input flooring (epsilon = 1e-8) ensures 100% finite arithmetic, but improper flooring on near-zero inputs causes significant bias, while output capping (L = 30) prevents floating-point overflow without distorting ordinary inputs. - Objective Matching: Analytic log-OLS serves as an exact same-family reference (WAPE = 0.0666 on unequal powers), while 2,000-update Adam fits demonstrate remaining optimization gaps (WAPE = 0.3443). How to Use the Data: Run "python analysis/replay.py all --output ../replay_out" to verify saved predictions and metrics, or execute "python analysis/figures/reproduce.py" to regenerate all manuscript figures. Full SHA-256 checksums and data maps are provided in DATA_INDEX.md, REPRODUCE.md, and MANIFEST_SHA256.csv.
Files
Steps to reproduce
Software Environment & Dependencies: Experiments are conducted in Python 3.12 using NumPy 2.x and PyTorch 2.x running on a single-thread FP64 CPU backend. Full environment specifications, exact package dependencies, and SHA-256 file manifests are detailed in REPRODUCE.md and MANIFEST_SHA256.csv. Data Generation Protocols: 1. Primary Power-Law Grid: Generated via NumPy default_rng(910000 + 1000*s + 10*f + n) across four power target families (product, ratio, unequal powers, shared power), sample sizes n in {32, 64, 256}, log-noise SD sigma in {0, 0.01, 0.05, 0.1, 0.2, 0.5}, and 24 paired seed clusters (s=0..23). In-distribution (ID) log-inputs are drawn uniformly from [-1, 1]; out-of-domain (OOD) log-inputs are sampled from U[1.2, 2.2] in magnitude with random signs. 2. Non-Power & Diagnostic Grid: Targets additive, mixed, and smooth non-power targets across n in {64, 256}, sigma in {0, 0.1, 0.5}, and 24 seeds. 3. Observational Data: Grouped 22-hull out-of-fold evaluations for Yacht Hydrodynamics (14 Froude-speed observations per hull) and 22 distance-based folds for NIST Chwirut2. Optimization & Experimental Protocols: 1. Primary L-BFGS-B Fits: Analytic-gradient L-BFGS-B with max 300 iterations, function tolerance 1e-12, and gradient tolerance 1e-8. 2. Matched-Budget Adam Controls: 2,000 full-batch Adam updates per model across 2 independent random initializations. Model selection uses an independent 128-observation noisy ID validation set generated with seed (620261019 + 1000*f + s), retaining the best validation checkpoint without observing OOD data. 3. Guard Sweep & Mixture Audits: 864 OLS refits crossing epsilon in {1e-12, 1e-8, 1e-6, 1e-4}, L in {2..50}, and FP16/32/64 precision modes; 1,152 mixture fits crossing alpha in {0, 0.1, 0.25, 0.5, 0.75, 1.0}. Execution & Reproduction Workflow: 1. Verify Data Integrity: Confirm SHA-256 checksums against MANIFEST_SHA256.csv. 2. Replay Saved Predictions: Execute "python analysis/replay.py all --output NEW_DIRECTORY" to re-verify stored predictions and metrics against saved model checkpoints. 3. Re-train Baseline Models: Navigate to "experiments/fair_baselines/portable/" and run "python portable_entry.py smoke --output NEW_DIR" (for quick verification) or "python portable_entry.py rerun --output NEW_DIR" (to re-fit all 2,520 Adam models). 4. Figure & Table Regeneration: Run "python analysis/figures/reproduce.py --output NEW_DIR" to re-generate all manuscript figures directly from verified metric CSV files.
Institutions
- Tzu Chi UniversityTaiwan, Hualien City
- National Pingtung UniversityTaiwan, Pingtung City
Categories
Funders
- Tzu Chi FoundationTaiwan, Hualien CityGrant ID: TCMMP114-01, TCAS111-02, TCAS112-02, TCAS113-04, TCRD112-033, TCRD113-041, TCRD114-029, and TCRD115-030