Reproducible data examples¶

These five downloadable examples use the same construction and analysis workflow with different variables, objectives and neighborhoods. Their recorded results are computed by GraphFLA. They are separate from the homepage's illustrative demo tables, authored landscapes and example profiles.

Example	Analyzed input	Objective	Neighborhood
Protein engineering	16 measured GB1 variants, using C/V, D/E, A/G and A/V at the four variable sites	Maximize binding strength	One amino-acid substitution
Chemistry	384 Suzuki–Miyaura conditions	Maximize UV product-area percentage	Replace one ligand, base or solvent
Materials	496 W–Re–Os compositions	Maximize hardness at 1,000 °C	One adjacent W or Re level; Os is the remainder
Software tuning	64 measured LLVM configurations; six flags vary and four stay off	Minimize compilation time	Flip one flag
Hyperparameters	48 random-forest settings, evaluated on the built-in scikit-learn diabetes regression dataset	Minimize validation RMSE	One adjacent depth, leaf-size or feature-sampling level

The HPO grid varies maximum depth (2, 4, 6, 10), minimum leaf size (1, 2, 4, 8) and feature fraction (0.5, 0.75, 1.0). Each forest has 40 trees and uses one CPU thread. A fixed 75/25 train/validation split and model seed 42 make the comparison reproducible. These are measured model evaluations, not invented scores.

Interpreting the four results¶

Local optima counts configurations or neutral plateaus with no improving move. The protein example uses roughness-to-slope ratio; the other examples use lag-one autocorrelation, estimated from 200 walks of length 20 with seed 42.

Near-equal neighbors uses an absolute tolerance equal to 1% of that example's observed objective range. The displayed tolerance is in the original objective units. This does not claim exact equality or experimental insignificance.

Interaction patterns shows exact directed four-node motif proportions from classify_epistasis: magnitude, sign and reciprocal-sign classes, labeled by whether local changes retain their direction, reverse one effect, or reverse both. Missing or equal-outcome edges can exclude a motif. These proportions describe the constructed graph, not every possible combination of variables.

Recorded graph coordinates¶

The recorded coordinates describe at most 32 connected configurations selected from the analyzed graph, including its best observed configuration. Edges are actual neighbor pairs. The planar arrangement is a force-directed projection, not a physical distance. Objective values are normalized so better outcomes are higher even for minimization. The reported metrics use the full analyzed input in the table above.

Download and reproduce¶

The repository script docs/home/prepare_scenarios.py reproduces these examples in an environment with GraphFLA, pandas, scikit-learn and igraph installed. Normal documentation builds render the recorded data without fitting models or running analysis. Original experimental sources are described in the dataset collection and worked tutorials.