Using the parameter-importance chart

Every successful submission result page renders a Parameter importance chart. Each bar shows the Pearson correlation between one gene's value across the last generation and the corresponding fitness values. The chart works the same way for single- and multi-objective runs; in the multi-objective case you see one bar per objective per gene, grouped side-by-side.

What the colours mean¶

• Positive (success-coloured) — increasing this gene tends to grow fitness on this run.
• Negative (danger-coloured) — increasing this gene tends to shrink fitness.
• Near-zero (very short bar) — this gene is not driving the outcome in the final-generation population. Either the algorithm has already pinned it to a winning value (it stopped varying), or it's genuinely orthogonal to the objective.

The bars are sorted by magnitude so the most influential genes appear at the top — across any direction, across any objective.

A practical workflow¶

1. Run the problem once with whatever search space feels reasonable.
2. Open the result page and look at the parameter-importance chart.
3. Identify any gene whose bar is essentially flat across every objective. Those are candidates to pin in the next run: replace the random range with a single value (or a tight ±5 % around the value the chart converged to) so the algorithm spends its budget on genes that actually matter.
4. Resubmit with the narrowed search space. The fitness should reach the same plateau in fewer generations.

A worked example¶

Imagine a 6-gene single-objective run where the parameter-importance chart says:

Gene 0   ████████████████   +0.82
Gene 3   ████████████       +0.61
Gene 5   ███                -0.18
Gene 1   ▏                   0.02
Gene 2   ▏                  -0.01
Gene 4   ▏                  -0.03

Genes 1, 2, and 4 are noise — they're not pulling the fitness in any direction. Set their init_range_low / init_range_high to the same value the best solution converged to, and rerun. With those three genes fixed, the algorithm has effectively a 3-dimensional search instead of 6-dimensional, and you'll reach the optimum in a fraction of the generations.

What the chart cannot tell you¶

Pearson correlation is linear. It will not detect a relationship that is U-shaped or that emerges only inside a narrow band. If a gene has a flat bar but you have a domain reason to believe it matters, do not prune it.

Related¶

• Tuning the mutation rate — once you've identified important genes, focus mutation there.
• Reading a result — full result-page tour.