Comparing two runs side by side

When you change a single GA parameter — say widen the mutation range or raise sol_per_pop — the natural question is "did that actually help?". Eyeballing two result pages in different tabs is error-prone. Vilvik has a dedicated compare view that lines two (or more) runs up so the deltas are obvious.

When to reach for compare¶

• You're tuning a hyper-parameter and want to know which value the chart converges on fastest.
• You're explaining a result to a teammate and need a single screenshot of the trade-off.
• You suspect a regression — a previous run produced a better fitness with apparently fewer resources.

The flow¶

1. Run the baseline submission. Note the result's URL or just leave the tab open.
2. Run the second submission with exactly one parameter changed. Keep everything else identical so the comparison stays clean.
3. Open the compare page and click Add another run to pick a second result.

The compare page renders three side-by-side panels:

• Fitness chart overlay — both runs' best-fitness curves on the same axes. The faster a curve climbs and the higher its plateau, the better the parameter setting.
• Parameter table — every parameter, with the rows where the two runs differ highlighted. Useful when you forgot which knob you actually changed.
• Genome comparison — the best solution from each run. For continuous problems you'll often see they converge on the same neighbourhood; for combinatorial problems you'll see different orderings.

Reading the overlay¶

Two healthy patterns:

• Same plateau, faster climb — the change made the algorithm converge in fewer generations. Cheap win; you saved CPU seconds.
• Higher plateau — the change found a better optimum. Bigger win; you found a previously inaccessible region of the search space.

Two anti-patterns:

• Same plateau, slower climb — the change is making the algorithm work harder for nothing. Roll back.
• Lower plateau — the change hurt. Roll back and try the opposite direction.

Comparing more than two runs¶

Add as many runs as you like; the compare page handles n runs. Charts get busy past ~4 series, so consider deleting earlier runs from the picker once you've sketched the trend.

Related¶

• Compare two runs — compare page reference.
• Tuning the mutation rate — the most common reason to open compare in the first place.