Skip to content

Relative-score estimation: contest-style best/your normalization with a maintained best-known reference #11

Description

@iray-tno

(Extends #10's relative metric into the actual scoring used by many optimization contests.)

Many optimization contests don't score on absolute value but on a relative per-case formula — e.g. for minimization score_i = round(K · best_i / your_i) summed over cases, where best_i is the best result across all competitors. The key properties: each case is worth the same maximum, and a fixed % improvement on an already-good case earns the same as on a weak one — so the plain absolute sum is a misleading optimization target.

You can't see competitors' global best_i, so it must be estimated from a maintained reference. Proposed:

  1. Persistent best-known-per-case store — updated automatically as you run configs/the suite over time (and optionally seeded from / merged with a user-provided best-known file). This is the reference best_i.
  2. Relative-score estimate report — in the contest's units (Σ round(K · best_i / your_i)), plus % of theoretical max and an estimated standing, computed against the maintained reference.
  3. Per-case relative contribution — rank cases by how many relative points they cost (1 − best_i/your_i), which surfaces the cases worth optimizing — very different from the absolute-gap ranking, since absolute gaps over-weight large-magnitude cases.

This turns the suite runner (#10) from "absolute aggregate" into a faithful estimate of relative standing, which is what actually determines rank in relative-scored contests, and lets the optimizer (#9 / Optuna) target the right objective.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions