Log — Optimization Algorithms Wiki

Append-only history. Entries start with ## [YYYY-MM-DD] <op> | <title>.

[2026-06-09] split | optimization-algorithms-wiki created from _inbox (4 sources)

Spun out of the hub optimization-algorithms _inbox cluster once it blew past 3 — the user dropped four entries from Andrey Dik's MQL5 optimizer-benchmark series in quick succession: exchange-market-algorithm (EMA), backtracking-search-algorithm (BSA), cma-es, and deterministic-oscillatory-search (DOS). EMA & BSA had been parked in _inbox (cluster 1→2); CMA-ES completed the cluster and DOS arrived during the build. Scaffolded raw/+wiki/, CLAUDE.md from the hub template, and these spine files; added the registry block to ../wikis.md; bumped the hub spoke count 12 → 13. Founding page set: 4 source/algorithm pages + concepts (metaheuristic-optimization, exploration-vs-exploitation, population-optimization-benchmark, no-free-lunch-theorem) + andrey-dik. All sources URL-only (raw/ empty).

[2026-06-09] ingest | founding sources (EMA, BSA, CMA-ES, DOS — MQL5 series)

Ingested the four cluster sources per Ingest. See synthesis: these are all population-based metaheuristics scored on one comparative benchmark (population-optimization-benchmark), and the headline is No Free Lunch — even the celebrated cma-es only ranked 38/45 on this suite, while simple ideas varied wildly (BSA 20/45; EMA & DOS near the bottom). Axes that emerge: stochastic-vs-deterministic (deterministic-oscillatory-search is fully reproducible), scalability (cma-es is O(n³), fails at high dimensions), and the universal exploration-vs-exploitation balance.

[2026-06-09] ingest | +5 classics from the Dik series (PSO, GWO, ACO, ABC, ES)

Expanded the spoke by five more entries from andrey-dik's MQL5 series, adding the canonical swarm/EA classics the founding four lacked: particle-swarm-optimization (PSO, Kennedy & Eberhart 1995), grey-wolf-optimizer (GWO), ant-colony-optimization (ACO, Dorigo — recast for continuous spaces), artificial-bee-colony (ABC, Karaboga 2005), and evolution-strategies ((μ,λ)-ES & (μ+λ)-ES, Rechenberg/Schwefel). Spoke now 9 algorithms / 14 pages. Two findings sharpen the no-free-lunch-theorem thesis (see synthesis): (μ+λ)-ES leads the entire mature suite at 72.18% — a 1970s strategy beating every modern method, with the comma/plus survival rule alone swinging rank ~20 points — and PSO ranked below the random-search baseline in Dik's early scoring. Also surfaced a benchmark-comparability caveat: Dik's scoring evolved (early small-field 0–1 over Skin/Forest/Megacity vs. mature %-of-MAX over Hilly/Forest/Megacity), so PSO/GWO/ACO/ABC numbers are not comparable to the founding-four percentages — population-optimization-benchmark now keeps the two schemes in separate tables.

[2026-06-09] ingest | +5 independent authoritative sources (Wikipedia) — break single-author dependence

At the user's request, expanded with 5 non-Dik, authoritative (Wikipedia) sources so the thesis no longer rests on one MQL5 series: differential-evolution (Storn & Price), genetic-algorithm (Holland — family ancestor), simulated-annealing (Kirkpatrick 1983 — the corpus's first single-solution / trajectory method), test-functions-for-optimization (the independent academic benchmark basis — Rastrigin/Rosenbrock/Ackley behind CEC/BBOB), and bayesian-optimization (the corpus's first model-based / surrogate optimizer). Spoke now 14 algorithms/concepts / 19 wiki pages. Effects on the thesis (see synthesis): (a) two new axes — population-vs-single-solution and model-free-vs-model-based; (b) independent corroboration of no-free-lunch-theorem (GA's article concedes SA/hill-climbing "often outperform" it; SA's convergence guarantee is practically useless); (c) the CMA-ES Dik-vs-BBOB tension promoted from open question to a recorded contradiction (≈38/45 on Dik vs. near-top on BBOB — kept both, flagged, not overwritten). Refreshed no-free-lunch-theorem, metaheuristic-optimization, and population-optimization-benchmark in place with the independent grounding.

[2026-06-09] ingest | +3 the non-metaheuristic half (gradient descent, SGD/Adam, convex) — all-spokes cron test

Closed the "where do gradient-based / exact methods fit?" open question with three Wikipedia-sourced foundations: gradient-descent (DefinedTerm, src — first-order, derivative-using, local), stochastic-gradient-descent (DefinedTerm, src — SGD + momentum/AdaGrad/RMSProp/Adam; the optimizer ML training runs on; bridge to llm-providers/llm-inference), convex-optimization (DefinedTerm, src — local=global, polynomial-time; the structure⇄generality trade). Reframed the whole wiki: founding corpus = the gradient-free/black-box/ no-guarantee quadrant; these are its complements. Synthesis open question answered + unifying "structure ⇄ generality" frame made explicit (an NFL restatement). 19 → 22 pages.

[2026-06-10] ingest | Nelder–Mead + Tabu search — all-spokes pass (two classic missing quadrants)

Two canonical methods filling gaps in the gradient/exact/metaheuristic map (spoke heavily grown 06-09, so kept to 2 high-value classics). nelder-mead (DefinedTerm, source, Wikipedia) — Nelder & Mead 1965 downhill-simplex; derivative-free direct search that is local + (largely) deterministic + single-solution (n+1 simplex via reflection/expansion/contraction/shrink) — a quadrant neither the global-stochastic metaheuristics nor the gradient/exact methods occupied. tabu-search (DefinedTerm, source, Wikipedia) — Glover 1986; memory-based single-solution metaheuristic completing the single-solution axis beside simulated-annealing (both accept worsening moves; SA=stochastic cooling vs TS=deterministic tabu-list memory), strong on combinatorial/discrete problems the continuous population-optimization-benchmark doesn't test. Both reinforce no-free-lunch-theorem by niche. Folded into synthesis (new 2026-06-10 section) + index (2 DefinedTerm rows). No contradictions. 22 → 24 pages.

[2026-06-12] ingest | COCO / BBOB benchmarking platform — numbbo

All-spokes daily expansion. Added coco-bbob (@type SoftwareApplication) — the neutral academic benchmarking platform the "how neutral is the benchmark?" open question wanted. COCO ("Comparing Continuous Optimisers") runs the GECCO BBOB workshops (2009–present): standard suites (bbob, bbob-largescale, bbob-biobj, mixed-integer), fixed-target/ERT methodology, multi-language bindings, automated experiment→figure pipeline — reproducible and cross-comparable, the counterweight to one author's MQL5 suite. Gives the recorded CMA-ES contradiction (~38/45 Dik vs near-top BBOB) a concrete academic home without overturning Dik (record-don't-overwrite); also covers discrete/mixed-integer regimes the continuous-only MQL5 suite omits. Open question reframed (platform now paged; literal Dik-in-COCO cross-run still pending). Wired to test-functions-for-optimization / cma-es / no-free-lunch-theorem; synthesis (open Q + contradiction) + index updated. 1 new page. Authoritative (academic standard platform).

[2026-06-15] ingest | NFL primary paper (Wolpert & Macready 1997) — T1 anchor for the central theorem

Quality cycle, T1-floor raise. The spoke's organizing claim (no-free-lunch-theorem) was grounded only in Andrey Dik's MQL5 blog (T3) + Wikipedia echoes; added the canonical primary as a source page: nfl-original-paper (ScholarlyArticle, T1) — Wolpert & Macready, No Free Lunch Theorems for Optimization, IEEE TEC 1(1):67–82, 1997. Captures Theorem 1 (∑_f P(d^y_m|f,m,a) identical across algorithms), the cost-value-sequence framework, the time-varying NFL, the alignment-with-P(f) / geometric reading, and head-to-head minimax. Separates theorem (T1 primary) from empirical illustration (Dik suite, T3). Sourcing note: found via WebSearch; the IEEE PDF didn't machine-extract (no pdftotext/pypdf in env), so formal statements were corroborated against the Wikipedia NFL article and the well-established canonical formulation — recorded transparently on the page. Linked from the concept page, synthesis ("Independent corroboration"), and index (new ScholarlyArticle section). 1 new page.

[2026-06-18] ingest | DefinedTerm enrichment pass (subagent)

Thinnest-first deepening of DefinedTerm pages with new authoritative-sourced facts; 7 source summaries created (all T2 TechArticle, Wikipedia). Pages deepened: convex-optimization (standard form, LP⊂QP⊂SOCP⊂SDP ladder, KKT/complementary-slackness, Nesterov–Nemirovskii 1994 self-concordant barriers → wikipedia-convex-optimization); gradient-descent (Cauchy 1847, O(1/k)→O(1/k²) Nesterov, condition-number zig-zag, Barzilai–Borwein/momentum → wikipedia-gradient-descent); simulated-annealing (Metropolis-1953 genealogy, Černý-1985 co-credit, adaptive SA + restarts → wikipedia-simulated-annealing); stochastic-gradient-descent (Robbins–Monro 1951, learning-rate decay convergence, Adam=Kingma–Ba-2014, AdamW weight-decay decoupling → wikipedia-stochastic-gradient-descent); artificial-bee-colony (Karaboga-2005 canonical four phases, vᵢₖ=xᵢₖ+Φ(xᵢₖ−xⱼₖ), roulette onlookers, the limit abandonment knob → wikipedia-artificial-bee-colony); ant-colony-optimization (Dorigo's Ant System, (τ^α)(η^β) transition + ρ evaporation, MMAS/ACS, combinatorial-native vs Dik's continuous recast → wikipedia-ant-colony-optimization); grey-wolf-optimizer (metaphor-inflation: Sörensen critique, Glover-1986 term → wikipedia-metaheuristic). All 7 source pages added to index.md (new TechArticle section). New cross-links wired (e.g. ABC↔DE, ACO↔tabu-search, GWO↔ES). No fabrication — every added claim traces to a fetched source; no build/verify, no git. 7 source pages, 7 pages deepened.

[2026-07-25] ingest | Dingo Optimization Algorithm Modification (DOAm) — MQL5, Andrey Dik

Routed from the hub (route entry in ../log.md; runner-up: none — a general optimizer that merely ships in MQL5, which the spoke's boundary rule sends here). URL-only, T3 (practitioner benchmark, and see the neutrality note below). New page: dingo-optimization-algorithm (DefinedTerm, source: true) — the four-behaviour base DOA (group attack / persecution / scavenging / survival), Dik's three-edit DOAm, and the benchmark result. Dedup: grep for dingo|doa across wiki/ → no hits. No new entities (andrey-dik already paged; MQL5 not paged as an org, consistent with prior ingests in this spoke). Leader change — recorded, not overwritten. DOAm scores 73.63% and takes the top of the population-optimization-benchmark, 1.45 points above (μ+λ)-ES (72.18%). Both numbers kept: benchmark table now shows DOAm 1st and ES 2nd "leader until 2026-07"; evolution-strategies rewritten to say it led rather than leads, with the date range. Gap-relevance: hits two open questions at once. (1) Metaphor vs mechanism — DOAm beats a 1970s ES by flipping a sign in the group-attack move, dropping the survival threshold 0.3 → 0.01, and deleting an abs() from scavenging. The dingo story passes through the modification untouched; only arithmetic moved the score. Strongest instance yet of the Sörensen "metaphor ≠ novelty" critique the spoke carries via wikipedia-metaheuristic. (2) Benchmark neutrality — the suite is now led by the suite author's own modification of a published algorithm, unreplicated elsewhere. Not an accusation (code published, weaknesses stated), but the open question now has a concrete target: run DOAm on coco-bbob. Recorded on the page and in synthesis. Also noted: DOAm's own profile (Forest 0.912 / Megacity 0.848 vs Hilly 0.454, high variance, sticks on smooth) plus Dik's advice to use it as an exploration front-end and hand off — no-free-lunch-theorem restated from the top of the table, and exploration-vs-exploitation resolved by switching algorithms rather than tuning one. Verify deferred per hub policy (content-only, no page moves). avoid-ai-writing run.

[2026-07-26] ingest | algorithmsbook.com — Kochenderfer's three MIT Press Algorithms books

The spoke's first academic textbook source and its first T1 reference that isn't a paper or a Wikipedia article. In scope: Algorithms for Optimization (Kochenderfer & Wheeler, MIT Press 2019, second-edition preview up, free CC BY-NC-ND PDF, Julia implementations, notebooks on GitHub). New: algorithmsbook (source), mykel-kochenderfer (entity). Updated: bayesian-optimization (textbooks file it under surrogate models, not metaheuristics), metaheuristic-optimization (related), synthesis + one new open question. Synthesis: the book doesn't contradict the corpus, it re-proportions it — in the Dik series metaheuristics are the field; in the textbook "population methods" is one chapter among local descent, stochastic methods, linear/constrained programming and surrogate models. The structure ⇄ generality map survives that reframing, which is mild corroboration it was drawn right. New open question: what happens to the ranking question when the objective is a vector? Every benchmark here scores one number per run; the book's multi-objective chapter has no total order to rank against. Same for optimization under uncertainty. Neither has a page or a source. Landing page weakness recorded on the page: algorithmsbook.com carries no descriptive text at all — covers and author names only, so everything substantive came from the book's sub-page. Out of scope, noted not split: Algorithms for Decision Making (MDP/POMDP/RL) and Algorithms for Validation (unpublished) — same author, same free-PDF model, no spoke owns that body. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-09] ingest | Linear and integer programming, MIPLIB, and Mittelmann's benchmarks (via research pass)

Coverage edge 4, mostly closed. New pages: linear-programming and integer-programming (T2 — Wikipedia as a tertiary survey, matching this spoke's existing wikipedia-* practice), miplib (MIPLIB 2017 at ZIB, T1) and mittelmann-benchmarks (Hans Mittelmann, ASU, T2).

Why the gap mattered. Thirty-six pages of metaheuristics and nothing on the methods industry actually runs. Adding them changes how the spoke reads its own material: LP gives a global optimum, a polynomial-time guarantee and a duality certificate; branch and bound gives a solution plus a bound on how far from optimal it is. Every metaheuristic here returns a point and no idea how good it is. That is the missing half of the no-free-lunch-theorem argument — where structure exists, exploiting it buys guarantees no black-box search can match.

One finding was not about algorithms at all. Mittelmann's page documents its own dismantling: "Through an action by Gurobi at the 2018 INFORMS Annual Meeting this has come to an end. IBM and FICO demanded that results for their solvers be removed," with Gurobi and MindOpt withdrawing in 2024. The field's longest-running independent solver comparison is incomplete because of licence terms, not methodology. Recorded in synthesis beside the spoke's other benchmark critiques, which have all been methodological.

What did not close, stated plainly. The edge asked for a text chapter, and no free T1 text was reachable — the Kochenderfer PDF now redirects to Google Drive, Bixby's history of LP/MIP computation is paywalled. So two concept pages rest on Wikipedia, which is weaker than this spoke's better pages. No solver has a page either: HiGHS, CBC, SCIP and Gurobi are still unwritten. Both are named in the edge's successor.

Entities: 0 created.

[2026-08-10] quality | Cycle: the spoke's first solver, and a freshness regrade

Quick win #6 closed. highs written from the project's own docs (T1) and placed against mittelmann-benchmarks' current tables (LPopt 2026-07-01, MILP 2026-07-07, both read in full and folded back into that page). HiGHS is the strongest fully-open solver in both — and 12.8× / 7.55× off the COPT baseline, with 82 of 240 MIPLIB instances unsolved in two hours. Recorded as a lower bound on the real gap, since Gurobi, CPLEX and Xpress are absent from Mittelmann at their vendors' demand.

Freshness: 10 stale → 0, by regrade rather than refresh. Seventeen pages were marked freshness: volatile; ten had passed the 60-day window. Every one was an algorithm concept page or a summary of a dated MQL5/Wikipedia article — the exact case QUALITY.md says to grade stable, because re-reading the URL returns the same text and the flag can never be cleared by work. Regraded to stable; coco-bbob kept volatile as a live benchmark platform. No page's content was touched.

[2026-08-14] quality | research pass: the instrument for a vector objective

Growth edge 3 ("ranking when the objective is a vector") hunted. hypervolume-indicator (Guerreiro, Fonseca & Paquete, ACM Computing Surveys 54(6) 2021, T1) supplies the standard Pareto-front quality metric — strictly monotonic with dominance, #P-hard to compute exactly above ~4–5 objectives. Closes the instrument gap, not the edge: no page here has run a multi-objective benchmark yet. synthesis.md open question + growth edge 3 updated; index.md gains one ScholarlyArticle entry. Entities: 0 created.

[2026-08-22] quality | Cycle: the two primary papers the spoke was missing

Coverage edges 5 and 7 closed with the papers themselves, both T1. nsga-ii (Deb, Pratap, Agarwal & Meyarivan, IEEE TEC 6(2):182–197, 2002; read from the full PDF via cse.unr.edu, IEEE being paywalled) gives the multi-objective machinery: fast nondominated sorting at O(MN²) against NSGA's O(MN³), crowding distance in place of the sharing parameter, elitism via the combined P∪Q pool, and constrained domination that handles constraints without a penalty weight. Measured on nine problems (SCH, FON, POL, KUR, ZDT1–4, ZDT6) at 25,000 evaluations against PAES and SPEA.

adam-optimizer (Kingma & Ba, ICLR 2015) replaces the second-hand treatment the spoke had via wikipedia-stochastic-gradient-descent — the same "an encyclopedia entry is not the primary text" upgrade the hub has been tracking. Carries the bounded-step / trust-region argument, the exact AdaGrad and RMSProp reductions, and the defaults (α=0.001, β₁=0.9, β₂=0.999, ε=10⁻⁸).

Both pages record that their comparisons were run by the authors of the method being compared, which makes three instances in this spoke with DOAm. index.md gains two ScholarlyArticle entries; synthesis.md gets edges 5 and 7 struck with successors and a new section on the author-scored pattern. Entities: 0 created.

[2026-09-03] ingest | Huangfu & Hall, "Parallelizing the dual revised simplex method"

Research pass, quality cycle, closing the chapter half of coverage edge 4. huangfu-hall-parallel-simplex (T1, Mathematical Programming Computation 10(1):119–142, 2018) is the peer-reviewed engine paper behind highs, read from the arXiv preprint PDF with pypdf.

Dedup caught a stale edge before it caught a duplicate. Edge 4 said "no solver has a page: HiGHS, CBC, SCIP and Gurobi are named nowhere." highs was written 2026-08-10. The edge had been wrong for three weeks — the fourth instance this week of a growth edge outliving the source that answered it, all four now recorded in the hub's ## Outstanding.

Numbers on a 30-problem reference set drawn from Mittelmann's list: PAMI on 8 cores 1.51× faster than serial hsol, hsol 2.29× faster than Clp 1.15, hsol's iteration speed 1.60× Clp on 0.70× the iterations; PAMI's mean parallel speedup 2.34×, twice as fast on more than 65% of the set; SIP 1.15× and complementary. The result worth keeping is the negative one the authors flag themselves: PAMI in serial is worse than hsol (0.64) and still wins on eight cores, because suboptimization wastes work per iteration and buys back more by overlapping them. A cutoff factor of 0.95 bounds the waste; below 0.9 the mean speedup degrades.

Disclosure recorded on the page. Huangfu's affiliation is FICO, and the abstract states the techniques went into FICO Xpress and became "the basis of promotional material by FICO." A conflicted author declaring the conflict in a journal that publishes artifacts is a third arrangement beside the hub's "nobody re-runs anybody's numbers" item — neither a vendor benchmark nor an independent re-run. Contrast dingo-optimization-algorithm, where no such statement exists.

Edge 4 closed both halves; successor 4a opened — branch-and-cut is still undescribed, so the LP half has a primary and the IP half rests on Wikipedia. avoid-ai-writing run.

[2026-10-05] ingest | AdamW: Decoupled Weight Decay Regularization (quality cycle)

adamw (T1, Loshchilov & Hutter, ICLR 2019, read from the arXiv PDF). Closes coverage edge 7's successor: the spoke held AdamW only through wikipedia-stochastic-gradient-descent. L2 is not weight decay under Adam; decoupled decay gave 15% relative test-error gain on two image datasets, by the authors' own runs. Linked from adam-optimizer. No entity pages: neither author is named elsewhere in the corpus.

[2026-10-05] quality | quality cycle

Scorecard in the hub quality-log.md.

[2026-10-07] quality | edge 6 closed — L-BFGS (Liu & Nocedal 1989)

New page l-bfgs (T1, read from the authors' PDF with pypdf). The spoke's first quasi-Newton source. Edge 6 struck in synthesis; successor 6a asks for an independent benchmark, since the 1989 tests are the authors' own. Linked from gradient-descent and index. No published: field: the paper gives the year (1989) and no day.