Guided Fast Local Search
Algorithms AIMech:GuidedFastLocalSearch DRAFT
A local-search algorithm that combines GLS feature-penalty updates with FLS activation-bit search, reactivating subneighborhoods associated with newly penalized features and retaining solutions according to the original objective.
Identity Notes & Concept Scope
The actual 2026-10-15 post-cohort349 collection at commit f4b274e9c75703c0f3051ab630b4f090272f056a covers 780 maintained and 442 generated records. Four fixed-string, word-boundary searches included ignored and hidden files across the original checkout and three retained discovery/recovery checkouts. All 16608 occurrences were reconstructed into 1095 distinct contents, freshly read by two attributed supporting readers; root read 18 candidate-focused contents, all nine normalized definitions/21 fields and all 13 matching frontier objects. Normalization covered 4858 identity-field occurrences. No normalized candidate collision was found for the name, ID or explicit GFLS source spellings. Existing Guided Local Search supplies feature penalties and Fast Local Search supplies active-subneighborhood bookkeeping; the selected record identifies their explicit penalty-triggered reactivation controller, not either component alone. Shared source URLs and historical exclusions do not establish an existing composite record. Reuse owner275 and preserve all 907 other main and 393 nested entries. This bounded result does not imply global absence or independent approval.
Claims & Grounded Evidence
5 Claims
Scope
#named-penalty-activation-controller
Guided Fast Local Search is the named combined controller that repeatedly applies FLS to a feature-penalized objective and couples selected penalty updates to reactivation of related subneighborhoods.
Mechanism
#initialization-and-current-objective
Initialize a constructed candidate, set every feature penalty p_i to zero, and set every activation bit to one. For binary feature indicators I_i(s), search using the current augmented objective h(s)=g(s)+lambda*sum_i p_i I_i(s). The penalties in h are the current penalty state, not fixed at their initial zero values.
Mechanism
#strict-active-subneighborhood-search
Within a FLS call, scan active subneighborhoods in the configured order and accept the first encountered move that strictly improves the current augmented objective. Keep the successful subneighborhood active, activate the configured subneighborhoods affected by the accepted move, update the candidate, and continue to the next subneighborhood instead of restarting the complete scan. Deactivate a subneighborhood when its scan finds no improving move, and return when all activation bits are zero.
Mechanism
#utility-penalty-and-reactivation
After FLS returns, compute feature utilities I_i(s)c_i/(1+p_i), increment every maximum-utility feature penalty by one, and activate the subneighborhoods returned by the feature-to-subneighborhood mapping before the next FLS call.
Scope
#original-objective-incumbent-and-stopping
The outer procedure retains the best solution found according to the original objective g and stops by a supplied stopping criterion; feature/move activation mappings depend on the optimization problem.
⚠️ Curated Limitations & Negative Results
- All inactive bits do not certify an exhaustive local minimum unless the move enumeration and reactivation policy establish the required coverage. Subneighborhoods may overlap, and the method can miss improvements.
- With nonnegative feature costs, a positive maximum utility selects present features, but when every utility is zero the printed all-maxima loop can also select absent features although the prose discusses present features. A boundary convention must be specified for implementation; none is silently added here.
- Problem-specific features, costs, move enumeration, scan order, move/feature activation callbacks, incumbent bookkeeping and stopping semantics are required. The mathematical objective must reflect current penalties, and bit changes must persist as intended across calls; pseudocode is not a language-level API specification.
- Penalty guidance is a soft objective change, not a forbidden move or a guarantee of escape. With lambda=0, or features constant over the relevant moves, penalties do not change those move comparisons. Strict inner descent does not by itself ensure the outer process terminates or finds a global optimum.
- The core FLS rule is strict improvement. Problem-specific variants such as the radio-link example allowing sideways moves and an extra stagnation stop should not be silently included in that core or inherit its conditional finite-descent argument.
- The sources motivate and report efficiency gains in selected applications, but the complete controller description does not establish a general runtime advantage, universal quality improvement, global convergence or an application-independent parameter prescription.
- The inspected evidence is the exact 1999 author manuscript and 2010 chapter layout. Their named complete procedure supports this scope; it does not establish earliest priority, equivalence of every later GLS/FLS hybrid, or exhaustion of later variants and neighboring search methods.
Artificial Intelligence Ontology (AIO) Grounding
Explicit AIO Snapshot Assessment: No Match
Rationale: No informative positive AIO mapping is supported for this complete source-defined search controller among the nine inspected targets. The result is bounded to the pinned snapshot and selected semantic neighborhood.
Search Scope: The actual 2026-10-15 post-cohort349 collection at commit f4b274e9c75703c0f3051ab630b4f090272f056a covers 780 maintained and 442 generated records. Four fixed-string, word-boundary searches included ignored and hidden files across the original checkout and three retained discovery/recovery checkouts. All 16608 occurrences were reconstructed into 1095 distinct contents, freshly read by two attributed supporting readers; root read 18 candidate-focused contents, all nine normalized definitions/21 fields and all 13 matching frontier objects. Normalization covered 4858 identity-field occurrences. The pinned AIO CLASS identities, synonyms and graph context for the nine explicitly listed examined targets were separately read and rejected for the maintained GFLS algorithm scope. General task/training/network/function abstractions and unrelated learning families do not cover its controller. No suitable target was identified in this inspected snapshot and semantic scope; this is not a claim about all upstream revisions.
| Examined AIO Candidate | Rejection Reason |
|---|---|
AIO:TrainingStrategy
|
The pinned definition concerns preprocessing used to train machine-learning models. This general combinatorial optimization controller is not intrinsically such a training method. |
AIO:MachineLearningTask
|
The pinned class requires methods learning from data. Updating search penalties and activation state for a supplied objective does not by itself establish that task identity. |
AIO:ReinforcementLearning
|
Search guidance and exploration do not alone establish the pinned machine-learning task. The source defines feature penalties and selective local search, not a reinforcement-learning task or value/policy-learning construction. |
AIO:ReinforcementLearningLLM
|
The candidate is not a large language model fine-tuned with rewards. |
AIO:EvolutionaryLLM
|
The candidate is neither a large language model nor a population-based evolutionary model architecture. |
AIO:MarkovChain
|
Representing an appropriately expanded algorithm state by a stochastic process is a generic formalization, not evidence that this whole source-defined controller is the pinned network/model class. |
AIO:MathematicalFunction
|
The pinned definition is a general mathematical input/output rule and is not scalar-only. The candidate is the complete configurable search controller, with objective, move, activation and stopping policies; generic functional describability alone does not establish an informative mapping to this class. Its augmented objective and utility formulas are mathematical components, not the whole candidate identity. |
AIO:RegularizationLayer
|
The feature-penalty objective operates on candidate solutions in combinatorial search, not penalties on neural layer parameters or activity. |
AIO:MetaLearning
|
The source does not define learning from metadata about machine-learning experiments. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
voudouris2010
|
Christos Voudouris, Edward P. K. Tsang, Abdullah Alsheddy
|
primary monograph
full_text
|
ID: doi:10.1007/978-1-4419-1665-5_11
sha256:e4dd9aed0fcf…
|
voudouris1999
|
Christos Voudouris, Edward Tsang
|
primary paper
full_text
|
ID: doi:10.1016/S0377-2217(98)00099-X
sha256:4550ebdad45d…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-15T20:34:02Z
Updated: 2026-10-15T20:34:02Z
Source File:
data/curated/GuidedFastLocalSearch.yaml