F-Race
Algorithms AIMech:FRace DRAFT
A statistical racing method for configuring metaheuristics that evaluates a finite candidate set on shared problem-instance blocks and eliminates candidates judged worse through a Friedman rank test followed by comparisons with the best-ranked candidate.
Identity Notes & Concept Scope
This record covers the original 2002 F-Race method. It uses a fixed starting candidate pool and nested survivor sets, not iterative resampling of configurations as in later iterated racing methods. It differs from ROAR’s challenger-incumbent heuristic comparisons, fixed-fraction successive halving and a standalone Friedman test without sequential resource allocation. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 265 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout, returning 20 lines in the worktree and 25 in main. These were prior candidate leads, source references, neighboring-method mentions and five unrelated ROAR namespace entries in ignored dependencies, not existing records for these identities. Seven exact AIO CLASS nodes with synonyms and outgoing edges and six neighboring native mechanisms were inspected. Absence is bounded to this snapshot and semantic neighborhood.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
Start with a finite set of candidate configurations and draw a sequence of training instances independently from the selected instance distribution. At each racing step, evaluate every surviving candidate on the same new instance and append its observed solution cost. Rank surviving candidates by cost within each accumulated instance block, using average ranks for ties. Apply the paper’s Friedman aggregate rank statistic to the accumulated blocks. If its null is not rejected, retain all survivors; if rejected, compare each candidate with the best-ranked one using the stated post-hoc rank-difference test and discard candidates found worse. Repeat with the reduced candidate set until one remains or the prescribed total computation budget stops further evaluation.
⚠️ Curated Limitations & Negative Results
- The method assumes independent instance blocks and appropriate within-block comparability. The aggregate test uses an approximate chi-square reference distribution, followed by the paper’s Student-t-based pairwise rule. No implementation-level small-sample correction or sequential family-wise error guarantee was verified.
- Rank comparisons discard cost magnitudes. Better rank sums need not identify the lowest arithmetic mean cost, so the method is not asserted to certify the minimizer of the expected-cost objective stated in Section2. Repeated significance decisions and early elimination do not ensure that the best configuration survives.
- A finite initial candidate pool limits what can be selected, and the time budget can expire with multiple survivors. The inspected method does not establish a unique returned configuration in every such case. No general superiority over parametric racing methods or replicated TSP result is claimed.
Artificial Intelligence Ontology (AIO) Grounding
Explicit AIO Snapshot Assessment: No Match
Rationale: No inspected AIO target captures this general algorithm-configuration procedure or a defining component. The local algorithm remains supported by its primary source; this is a bounded mapping assessment, not absence from all ontologies.
Search Scope: All 442 pinned AIO CLASS identities and scoped synonyms, all 265 preceding maintained identities, ignored/hidden checkout text, and semantic inspection of seven AIO targets plus neighboring configuration/resource-allocation methods.
| Examined AIO Candidate | Rejection Reason |
|---|---|
AIO:TrainingStrategy
|
The pinned definition concerns preprocessing/training of machine-learning models. This method configures target algorithms generally; an optional use for model tuning does not establish a training-strategy identity or necessary relation. |
AIO:MachineLearningTask
|
The target is a broad process for building methods that learn from data. Configuring an arbitrary solver over instances need not build such a learning method; shared use of observations does not justify equivalence or subclassing. |
AIO:MetaLearning
|
The pinned task learns from metadata about machine-learning experiments. These methods require no collection of ML experiments or learned cross-task prior; solver configuration alone is not that identity. |
AIO:ActiveLearning
|
The pinned task queries labels for data points. Choosing configuration evaluations or instance runs is not label acquisition in that defined sense. |
AIO:ReinforcementLearning
|
Exploration and empirical feedback do not by themselves establish the pinned learning-task identity; no reward-learning agent or policy-learning component is required by these configurators. |
AIO:RandomForest
|
These methods contain no defining decision-tree ensemble surrogate; random sampling, randomized local search and rank tests are not random forests. |
AIO:EvaluationBias
|
This is a bias concept, not a configuration algorithm. Potential overfitting or distribution mismatch is a limitation rather than an identity mapping. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Mauro Birattari, Thomas Stützle, Luis Paquete, Klaus Varrentrapp
|
primary paper
full_text
|
sha256:4e67ab87a580…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T12:04:49Z
Updated: 2026-10-11T12:04:49Z
Source File:
data/curated/FRace.yaml