You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Where the test problem library (docs/problems-plan.md, section 4) stands, batch by batch. I'll keep this issue up to date as batches land.
When a batch is done
A batch is done when every problem it adds has:
the problem itself in Rust and Python, with its tests and its citation in the module docs;
an example of its own in examples/<name>/, one problem per example: main.rs and main.py with the same output, a README (the problem, what makes it hard, representation, algorithm, output, good results), output.txt checked by CI, and the trace.json that its page on tachsin.gr/projects/genoxide/examples plays back, with a plot that shows the solutions;
a row in examples/README.md.
One problem per page: the comparison example function_suite was removed once each of its functions had its own page.
Where it stands
189 problems are in the library, and every one has its own example page (#253, #261, #270, #277, #307, #313, #357, #367, #401, #409, #414, #421). Every example reaches its problem's optimum or true front with its main method, or its README shows why it can't (#312; in batch 10a, Dixon-Price in 10-D, where CMA-ES and SHADE stall at a stationary point; in batch 10b, HappyCat and HGBat, whose minima neither CMA-ES with IPOP restarts over 10⁶ evaluations nor L-SHADE, the CEC 2014 winner, reaches; in batch 11, DAS-CMOP9, which MOEA/D-DE brings onto its front but whose 99% hypervolume target is beyond any 300 Tchebycheff weight vectors, as the README measures; and DC2/DC3-DTLZ, reached on the relaxed problem as C1-DTLZ3 is).
Batch 12: done in feat(problems): binary and combinatorial problems, batch 12, each with its own example #421. Each example reaches its exact optimum: NK landscapes by exhaustive search or dynamic programming, the knapsack by dynamic programming. It also found an erratum in Deb and Goldberg (1993): only Ackley's traps of 3 and 4 bits are fully deceptive, not all below 7.
Batch 13 (optional): competition suites with long definitions (LIR-CMOP, CEC 2009 UF/CF, MaF, and others), mainly for the benchmark suite.
Where the test problem library (docs/problems-plan.md, section 4) stands, batch by batch. I'll keep this issue up to date as batches land.
When a batch is done
A batch is done when every problem it adds has:
examples/<name>/, one problem per example:main.rsandmain.pywith the same output, a README (the problem, what makes it hard, representation, algorithm, output, good results),output.txtchecked by CI, and thetrace.jsonthat its page on tachsin.gr/projects/genoxide/examples plays back, with a plot that shows the solutions;examples/README.md.One problem per page: the comparison example
function_suitewas removed once each of its functions had its own page.Where it stands
189 problems are in the library, and every one has its own example page (#253, #261, #270, #277, #307, #313, #357, #367, #401, #409, #414, #421). Every example reaches its problem's optimum or true front with its main method, or its README shows why it can't (#312; in batch 10a, Dixon-Price in 10-D, where CMA-ES and SHADE stall at a stationary point; in batch 10b, HappyCat and HGBat, whose minima neither CMA-ES with IPOP restarts over 10⁶ evaluations nor L-SHADE, the CEC 2014 winner, reaches; in batch 11, DAS-CMOP9, which MOEA/D-DE brings onto its front but whose 99% hypervolume target is beyond any 300 Tchebycheff weight vectors, as the README measures; and DC2/DC3-DTLZ, reached on the relaxed problem as C1-DTLZ3 is).
bnh,kursawe)welded_beam,pressure_vessel,gear_train)Shifted<P>,Rotated<P>(17)Batch 9's tenth problem, conceptual marine design, was added in #401, completing batch 9.
Pages that were missing (batches 0 and 1), added in #261
The plan gives batch 10a a page per function (done in #367); batch 13 (benchmark suites) is to decide when it comes.
Next
Batch 10b follow-up (feat(problems): gradients for batch 10b's functions, and gradients and constraint values through Shifted and Rotated #416): done in feat(problems): gradients for batch 10b's functions, and gradients and constraint values through Shifted and Rotated #420. 14 of the functions supply analytic gradients; Step, the non-continuous Rastrigin, Katsuura and the noisy quartic don't, each documented.ShiftedandRotatedpass gradients and constraint values through.CTP8 against Deb's 2001 book: done in docs(ctp): CTP1-CTP8 checked against the published paper and Deb's 2001 book #419. All eight CTP problems match the published paper (EMO 2001) and the book (section 8.3.5); CTP8's parameters on p. 358 are the code's. Only CTP8's documented piece ends were slightly off, now corrected.Batch 12: done in feat(problems): binary and combinatorial problems, batch 12, each with its own example #421. Each example reaches its exact optimum: NK landscapes by exhaustive search or dynamic programming, the knapsack by dynamic programming. It also found an erratum in Deb and Goldberg (1993): only Ackley's traps of 3 and 4 bits are fully deceptive, not all below 7.Related
Deoptions), CarSideImpact's best known value can be beaten, and its docs misstate the active constraints #258 (car side impact's best known value), Ideal and nadir points for Poloni, Viennet 2, Viennet 3 and Kursawe #259 (ideal and nadir points).