Skip to content

Problems catalog

Choose a problem by its formulation, encoding and dimensions, then check the algorithm compatibility matrix. This catalog describes the 251 registered problems in VAMOS 1.4.0; Development after 1.4.0: CTP1–CTP7 bring them to 258, and IDTLZ1 and IDTLZ2 to 260. The named selection helpers are stable public APIs. A synthetic or engineering benchmark is a test model, not a validated model of your application.

Choose a starting point

What you want to exercise Start with Why
A first continuous run and a visible two-objective front zdt1 Inexpensive evaluation, 30 variables, packaged 2D reference sample
Three or more objectives dtlz2 Scalable objectives and a standard dimension rule
More varied geometry and difficulty wfg1–wfg9, zcat1–zcat20 Configurable objective count; record the family parameters
Hundreds of variables lsmop1–lsmop9 300 variables by default; group structure matters
Explicit constraint handling c1dtlz1, mw1 Emit inequality constraints separately from objectives
Binary, integer or mixed operators zdt5, int_alloc, mixed_design Small, deterministic examples with explicit encodings
Permutation operators tsp6, then kroa100 Closed tours with two specified route objectives
Application-inspired test functions RE, RWA, CRE and RCM tables below Fixed engineering formulations with different encodings and objective counts; CRE and RCM keep their constraints in G
Expensive model evaluations fs_real, ml_tuning Train a classifier for each candidate; require scikit-learn

Use optimize("zdt1", ...) in Python or --problem zdt1 in the CLI. Inspect the installed version through the public facade:

from vamos import available_problem_names, make_problem_selection

print(sorted(available_problem_names()))
selection = make_problem_selection("dtlz2", n_obj=3)
problem = selection.instantiate()
print(problem.n_var, problem.n_obj, problem.encoding)  # 12 3 real
print(problem.xl, problem.xu)

vamos.problems is a curated set of problem classes, not the catalog. It holds the members used most in examples and tests: ZDT1–ZDT4 and ZDT6 (pass n_var), DTLZ1–DTLZ4 and DTLZ7, WFG1–WFG9, CEC2009 UF1–UF3 and CF1, TSP, and three application examples (FeatureSelectionProblem, HyperparameterTuningProblem, and WeldedBeamDesignProblem). These classes are stable for 1.x. VAMOS 1.2.0: the classes added since cover a whole family and take the variant number and the family's parameters: DASCMOP, DCDTLZ, and LIRCMOP. Development after 1.4.0: CTP covers CTP1–CTP7 in the same way. Every registered problem, including the families and members without a class (ZDT5, DTLZ5 and DTLZ6, UF4–UF10, LZ09, MW, C-DTLZ, LSMOP, ZCAT, RE, and RWA; Development after 1.4.0: IDTLZ1 and IDTLZ2), is available by name through optimize("name") and make_problem_selection, the stable way to build any of them.

VAMOS 1.3.0: some registered problems also take options besides n_var and n_obj: capacity_ratio for bin_knapsack, and complicated_pareto_set, level, bias and imbalance for zcat1–zcat20, with the defaults in the tables below. Pass them by name with optimize(name, problem_kwargs={...}), a study's problem_kwargs, or make_problem_selection(name).with_options({...}); selection.spec.option_defaults() lists them. A run records them in resolved_spec.problem.config and replays them exactly. The other problems take no options, and arguments that identify a registered problem, such as a family's variant, stay fixed by its name.

The tables report the instantiated default dimensions. n is the number of decision variables, M the number of objectives, and G the number of explicit inequality outputs (G <= 0 is feasible). Zero G means no separate inequality outputs; bounds and encoding rules still apply. All objectives returned in F are minimized, including sign-flipped maximization objectives.

Reference sample means a CSV shipped with the package, with its objective dimension shown. It does not certify an exact Pareto front or guarantee suitability after changing the formulation. See Reference fronts before computing indicators.

Continuous test suites

These families use real-valued variables (real or continuous in the metadata) and need only the base installation, including WFG. Source links point to the corresponding implementation or family registration in the 1.0.1 release.

Registered keys Default n Default M G Default parameters / dimension rule Source Reference sample
zdt1, zdt2, zdt3 30 2 0 Variables in [0, 1]; fixed M ZDT 2D, each
zdt4 10 2 0 First variable [0, 1], remainder [-5, 5]; fixed M ZDT 2D
zdt6 10 2 0 Variables in [0, 1]; fixed M ZDT 2D
dtlz1 7 3 0 n = M + 4 (k = 5) when n is omitted DTLZ 2D and 3D; Development after 1.4.0: also 4D, 6D and 8D
dtlz2–dtlz6 12 3 0 n = M + 9 (k = 10); DTLZ4 uses alpha = 100 DTLZ 2D and 3D, each; Development after 1.4.0: DTLZ2–DTLZ4 also 4D, 6D and 8D
dtlz7 22 3 0 n = M + 19 (k = 20) DTLZ 2D and 3D; Development after 1.4.0: also 4D, 6D and 8D
idtlz1 7 3 0 Development after 1.4.0: n = M + 4 (k = 5) when n is omitted; DTLZ1's front turned over IDTLZ 2D and 3D
idtlz2 12 3 0 Development after 1.4.0: n = M + 9 (k = 10); DTLZ2's front turned over IDTLZ 2D and 3D
wfg1–wfg9 24 3 0 k = 4, l = 20; bounds [0, 2i] for variable i WFG 2D only, each
lz09_f1, lz09_f7, lz09_f8 10 2 0 Fixed M; family-specific linkage Li–Zhang LZ09 None packaged
lz09_f2–lz09_f5, lz09_f9 30 2 0 Fixed M; family-specific linkage Li–Zhang LZ09 None packaged
lz09_f6 10 3 0 Fixed M Li–Zhang LZ09 None packaged
cec2009_uf1–cec2009_uf7 30 2 0 Fixed M; UF5: N=10, UF6: N=2, both epsilon=0.1 CEC2009 2D, each
cec2009_uf8–cec2009_uf10 30 3 0 Fixed M; UF9: epsilon=0.1 CEC2009 3D, each
lsmop1–lsmop9 300 3 0 nk = 5 subcomponents per group LSMOP 2D only, each
zcat1–zcat20 30 2 0 level = 1; complicated Pareto set, bias and imbalance all False ZCAT 2D, 3D, 4D and 6D, each

Development after 1.4.0: IDTLZ1 and IDTLZ2 are the inverted DTLZ1 and DTLZ2 of Jain and Deb (2014), the usual test of how decomposition-based algorithms cope with an inverted front. They follow PlatEMO's IDTLZ1 and IDTLZ2: with DTLZ1's or DTLZ2's distance function g and position shape s (DTLZ1 is f = (1 + g) s / 2 and DTLZ2 f = (1 + g) s), IDTLZ1 is f = (1 + g) / 2 - (1 + g) s / 2 and IDTLZ2 f = (1 + g) - (1 + g) s. Every objective is minimized. With g = 0, at the distance variables 0.5, the front of IDTLZ1 is (1 - w) / 2 and that of IDTLZ2 1 - w / norm(w), for w on the unit simplex: DTLZ1's simplex and DTLZ2's sphere turned over, so that many of the evenly spread weight vectors of a decomposition point past it. With two objectives, IDTLZ1's front is DTLZ1's segment f1 + f2 = 0.5. pymoo 0.6.2's InvertedDTLZ1 computes IDTLZ1 in the same way; pymoo has no inverted DTLZ2. The classes are not exported by vamos.problems: build them by name. They take n_obj as DTLZ does.

DTLZ, WFG, LSMOP and ZCAT allow changing n_obj in named selection. For DTLZ, omitting n_var recalculates the standard dimension from M; setting a different n is allowed but changes the experimental setting. For WFG, k = 4 at M = 2, otherwise k = 2(M - 1), and l = n - k; WFG2/3 require an even distance-variable count. LSMOP needs enough variables for all groups to remain nonempty. Family constructor checks still apply to dimension overrides.

VAMOS 1.2.0: impossible dimensions are rejected with ValueError when the problem is built: WFG needs at least one distance variable (n >= k + 1, so wfg1 with M = 13 needs n_var >= 25), DTLZ needs n >= M, and LZ09 needs n >= 3. n_var and n_obj must be integers; a float or a boolean is rejected. VAMOS 1.1.0 accepted these settings and failed later with NaN objectives, an IndexError or a ZeroDivisionError.

VAMOS 1.2.0: lsmop1–lsmop9 follow the definition of Cheng, Jin and Olhofer (2017), as implemented in PlatEMO. For i >= M, LSMOP1–4 use the linear linkage (1 + i/n) x_i - 10 x_1 and LSMOP5–9 the nonlinear linkage (1 + cos(0.5 pi i/n)) x_i - 10 x_1. LSMOP9 divides each group's distance by sublen * nk, as LSMOP1–8 do. VAMOS 1.1.0 and earlier follow jMetal, which differs in two places: it uses the nonlinear linkage in all nine problems, and its LSMOP9 divides by nk only, so each group adds sublen times more to the distance (12 to 30 times at the default n = 300 and M = 3). The Pareto fronts and the packaged samples do not change, because every linkage term vanishes on the Pareto set. The Pareto set of LSMOP1–4 moves to x_i = 10 x_1 / (1 + i/n), and the landscape of LSMOP1–4 and LSMOP9 changes, so their results from those versions are not comparable with results on the original problems.

Explicitly constrained suites

All entries in this table are real-valued. They emit G alongside F; choose an algorithm with the required constraint support in the matrix.

Registered keys Default n Default M G Default parameters / objective count Source Reference sample
cec2009_cf1 10 2 1 Fixed M and constraint function CEC2009 None packaged
cec2009_cf2, cec2009_cf3 10 2 1 Fixed M; n ≥ 3 CEC2009 2D, each
cec2009_cf4, cec2009_cf5 10 2 1 Fixed M; n ≥ 2 CEC2009 2D, each
cec2009_cf6, cec2009_cf7 10 2 2 Fixed M; n ≥ 4 CEC2009 2D, each
cec2009_cf8–cec2009_cf10 10 3 1 Fixed M; n ≥ 5 CEC2009 3D, each
c1dtlz1 12 3 1 Configurable M; n remains 12 unless supplied C-DTLZ 2D and 3D
c1dtlz3 12 3 1 Default radius r = 9 at M = 3; varies with M C-DTLZ 2D and 3D
c2dtlz2 12 3 1 Default radius r = 0.4 at M = 3 and r = 0.5 otherwise C-DTLZ 2D and 3D
c3dtlz1 12 3 M Configurable M C-DTLZ 2D and 3D
c3dtlz4 12 3 M Configurable M; DTLZ4's alpha = 100 C-DTLZ 2D and 3D
dc1dtlz1, dc1dtlz3 12 3 1 Configurable M; a = 5, b = 0.95 DC-DTLZ 2D and 3D, each
dc2dtlz1, dc2dtlz3 12 3 2 Configurable M; a = 3, b = 0.9 DC-DTLZ 2D and 3D, each
dc3dtlz1, dc3dtlz3 12 3 M Configurable M; a = 5, b = 0.5 DC-DTLZ 2D and 3D, each
mw1, mw2, mw6, mw9 15 2 1 Actual M fixed at 2 MW 2D, each
mw3, mw7, mw12, mw13 15 2 2 Actual M fixed at 2 MW 2D, each
mw5, mw10 15 2 3 Actual M fixed at 2 MW 2D, each
mw11 15 2 4 Actual M fixed at 2 MW 2D
mw4, mw8, mw14 15 3 1 Configurable M MW 2D and 3D, each
lircmop1–lircmop12 30 2 2–3 Fixed M; variables in [0, 1]; LIR-CMOP1–4 need n ≥ 3 LIR-CMOP 2D, each
lircmop13, lircmop14 30 3 2–3 Fixed M LIR-CMOP 3D, each
dascmop1–dascmop6 30 2 11 Fixed M; jMetal's default difficulty DAS-CMOP 2D, each
dascmop7–dascmop9 30 3 7 Fixed M; jMetal's default difficulty DAS-CMOP 3D, each
binh2, constrex, srinivas, tanaka 2 2 2 Fixed n and M jMetal 2D, each
osyczka2 6 2 6 Fixed n and M jMetal 2D
golinski 7 2 11 Fixed n and M; Golinski's speed reducer jMetal 2D
water 3 5 7 Fixed n and M; water resource planning jMetal 5D
car_side_impact 7 3 10 Fixed n and M jMetal 3D
ctp1 5 2 2 Development after 1.4.0: fixed M; n ≥ 2; x1 in [0, 1], the others in [-5, 5]; the paper's J = 2 CTP 2D
ctp2–ctp7 5 2 1 Development after 1.4.0: fixed M; n ≥ 2; x1 in [0, 1], the others in [-5, 5]; the paper's (θ, a, b, c, d, e) CTP 2D, each
zxh_cf1–zxh_cf4 13 3 1–2 Configurable M; n = M + 10 unless supplied; type A ZXH-CF 3D, each
zxh_cf5–zxh_cf8 13 3 3–4 Configurable M; type B, k + 1 or k + 2 constraints ZXH-CF 3D, each
zxh_cf9–zxh_cf12 13 3 3–4 Configurable M; type C, k + 1 or k + 2 constraints ZXH-CF 3D, each
zxh_cf13–zxh_cf16 13 3 3–4 Configurable M; type D, k + 1 or k + 2 constraints ZXH-CF 3D, each

The constrained real-world suites CRE21–CRE51 and RCM01–RCM50 are listed with the application-inspired benchmarks.

DC-DTLZ's default constants are those of the authors' implementation (Li et al., 2019, in the COLA Laboratory's code), which pymoo also uses:

  • DC1: a = 5, b = 0.95;
  • DC2: a = 3, b = 0.9, with g / 100 in the constraints;
  • DC3: a = 5, b = 0.5;
  • DTLZ1's distance g, with factor 100, for both DTLZ1 and DTLZ3.

jMetal, PlatEMO and STEP (Cork et al., 2026) use a = 3, b = 0.5 and the unscaled g in the constraints of all six problems. PlatEMO also uses a factor of 10 in DC3-DTLZ3's g. Their problems therefore differ from these. Their DC1 and DC3 fronts differ too. Their DC2 fronts do not, because DC2's constraints hold on the whole DTLZ front.

The paper's supplementary document defines the six problems, and its text sets a = 3, b = 0.5 for DC1 and gives DC3 M + 1 constraints, on x1 to xM and on g. Its Table II, the share of uniform samples that is feasible, matches the authors' code instead: 10.11 % in DC1, which is arccos(0.95) / pi, against a third with b = 0.5; and 3.70 %, 0.41 % and 0.02 % in DC3 at M = 3, 5 and 8, which is (1/3)^M with M constraints, against (1/3)^(M + 1) with M + 1. The tests check VAMOS's defaults against that table. The text gives no constants for DC2 and DC3.

VAMOS 1.2.0: vamos.problems.DCDTLZ(key, n_var=..., n_obj=..., a=..., b=..., g_factor=..., constraint_g_scale=...) builds any of these variants:

  • a=3, b=0.5, constraint_g_scale=1 gives jMetal's and STEP's problems; the tests check them against jMetal.
  • g_factor multiplies DTLZ1's g, 100 by default.
  • constraint_g_scale scales the g that the constraints use: 0.01 in DC2 by default, and 1 in DC1 and DC3.
  • DC-DTLZ accepts as few as M − 1 variables, as STEP's instances with three objectives and two variables need; then g = 0.
  • DC3's constraint on g comes first, as in pymoo; jMetal puts it last.
  • The packaged fronts hold only for the default constants.

VAMOS 1.2.0: the C-DTLZ problems accept as many variables as objectives, with a single distance variable, as STEP's C3-DTLZ4 with three objectives and three variables needs. The tests check them against jMetal 7.10 with STEP's dimensions and, for C2-DTLZ2, its radius r = 0.1.

VAMOS 1.2.1: C2-DTLZ2's default radius is that of Jain and Deb (2014) and of jMetal's C2_DTLZ2: r = 0.4 with three objectives and r = 0.5 otherwise. VAMOS 1.2.0 and earlier used r = 0.2 with two objectives. With r = 0.5 the whole of the unconstrained front, the quarter of the unit circle, is feasible, and the packaged two-objective front is that quarter circle instead of the three arcs that r = 0.2 left. Pass r=0.2 to keep the former problem.

VAMOS 1.2.0: the classic constrained problems follow jMetal's classes of the same names, bounds and constraint scaling included. The front of Srinivas and Deb's problem is not only the textbook segment x1 = -2.5, which minimizes f1 + f2 and starts at f1 = 24.5. It goes on along the boundary 3 x2 - x1 = 10 down to the least feasible f1 = 10.1, at (1.1, 3.7).

VAMOS 1.2.0: LIR-CMOP1–14 (Fan et al., 2019) and DAS-CMOP1–9 (Fan et al., 2020) follow jMetal's implementation at commit 1e4b83cae6. For LIR-CMOP5–14 that is the paper's definition: jMetal corrected those problems in its pull request #522, after jMetal 7.8. DAS-CMOP's difficulty is the triplet (eta, zeta, gamma), each in [0, 1]:

  • eta controls diversity: how much of the front the first constraint cuts.
  • zeta controls feasibility: the width of the feasible band of the distance function. zeta = 1 turns it into an equality, met within 1e-4.
  • gamma controls convergence: the radius of the infeasible regions on the front.

The registered keys use jMetal's defaults: (0, 0.5, 0.5) for DAS-CMOP1 and 2, and (0.5, 0.5, 0.5) for the others. vamos.problems.DASCMOP(k, eta=..., zeta=..., gamma=...) builds another difficulty. The packaged fronts hold only for the defaults.

VAMOS 1.2.0: CEC2009 CF2–CF10 (Zhang et al., 2009) and ZXH-CF1–16 (Zhou, Xiang and He, 2021) follow jMetal's implementation at commit 1e4b83cae6; the reference values check them against it.

  • CEC2009 CF2–CF10 take jMetal's dimensions and bounds: n = 10 by default, the position variables (one with two objectives, two with three) in [0, 1] and the others in [-1, 1] (CF2), [-4, 4] (CF8) or [-2, 2]. G is the competition's c(x) >= 0 with the opposite sign. Where f3 = 1, CF8–CF10 divide by zero; as in jMetal, the constraint is then 0.
  • ZXH-CF1–16 have M = 3 and n = 13 by default. Another M gives n = M + 10 unless n is supplied. Variables are in [1e-10, 1 - 1e-10]. With S = x1^2 + ... + xM^2 and a distance function h (sphere, Rosenbrock, Ackley or Griewank, all 0 where the other variables are 0.2), every problem requires S + h <= 1, and the ones with a radius r of 1/4, 1/2 or 3/4 also S + h >= r. The first k angles of the front shape are then free (type A, CF1–4), in [1/4, 3/4] (type B, CF5–8), outside (1/4, 3/4) (type C, CF9–12), or outside (1/10, 2/5) and (7/10, 4/5) (type D, CF13–16). k is M − 1 up to three objectives, M // 2 up to eight, and then 3.
  • One constraint per angle: as the paper's k_theta constraints v_j(theta) <= 0 (its equation (6)), each constrained angle has one column: max(theta - 3/4, 1/4 - theta) for type B, min(theta - 1/4, 3/4 - theta) for type C, and the larger of min(theta - 1/10, 2/5 - theta) and min(theta - 7/10, 4/5 - theta) for type D. jMetal, whose classes are the authors' Java code, stores both bounds of each angle of type B and of each excluded interval of types C and D, and computes the violation and the feasibility of the intervals itself. The feasibility and the overall violation are the same, since an angle can break only one bound or fall in only one interval at a time; jMetal counts two or three violated constraints for an angle, VAMOS one.

VAMOS 1.2.1: ZXH-CF5–8 and ZXH-CF13–16 have one constraint per constrained angle, as ZXH-CF9–12 already had: k + 1 constraints, or k + 2 with a radius. VAMOS 1.2.0 stored two per angle, 2k + 1 or 2k + 2. The feasibility and the overall violation do not change; the number of constraints and of violated constraints does.

Development after 1.4.0: CTP1–CTP7 (Deb, Pratap and Meyarivan, 2001) follow the paper's definition. Both objectives are minimized, f1 = x1 with x1 in [0, 1], and each constraint c >= 0 of the paper is stored as G = -c:

  • CTP1 (the paper's eq. 4): f2 = g exp(-f1 / g), with J constraints f2 - a_j exp(-b_j f1) >= 0. The paper's procedure sets a_j and b_j so that each boundary meets the previous one, or the unconstrained front, at f1 = j / (J + 1). With J = 2 it gives a = (0.858266, 0.728234) and b = (0.541475, 0.295039), which the paper prints rounded to three digits.
  • CTP2–CTP7 (eq. 5): f2 = g (1 - f1 / g), with one constraint cos(θ) (f2 - e) - sin(θ) f1 >= a |sin(b π (sin(θ) (f2 - e) + cos(θ) f1)^c)|^d and the paper's parameters:
Problem θ a b c d e
CTP2 −0.2π 0.2 10 1 6 1
CTP3 −0.2π 0.1 10 1 0.5 1
CTP4 −0.2π 0.75 10 1 0.5 1
CTP5 −0.2π 0.1 10 2 0.5 1
CTP6 0.1π 40 0.5 1 2 −2
CTP7 −0.05π 40 5 1 6 0
  • The distance function g: the paper's experiments use "Rastrigin's function as the g functional" with five variables, without its formula. VAMOS uses the Rastrigin function of ZDT4 (Deb, 1999), g = 1 + 10 (n - 1) + Σ_{i=2..n} (x_i² - 10 cos(4π x_i)) with x_i in [-5, 5], whose minimum is g = 1 at x_i = 0, and n = 5 by default (n ≥ 2). It evaluates the same function as 1 + Σ (x_i² + 20 sin²(2π x_i)).
  • Other instances: vamos.problems.CTP(variant, n_var=..., n_constraints=..., theta=..., a=..., b=..., c=..., d=..., e=...) takes CTP1's J as n_constraints and the six parameters of CTP2–CTP7, with θ in radians. c must be a positive integer, since its base sin(θ) (f2 - e) + cos(θ) f1 can be negative. The registered keys take only n_var, and the packaged fronts hold only for the paper's parameters.
  • Other implementations: jMetal has no CTP problems. pymoo 0.6.2 writes the constraint of CTP2–CTP7 as the NSGA-II C code's ratio 1 - exp1 / exp2, where exp1 >= exp2 is the paper's constraint. The ratio is infinite where the sine is 0, and undefined where exp1 is 0 too; elsewhere it has the feasibility of the paper's difference exp1 - exp2, which VAMOS uses, but other values. pymoo's g also differs (1 + Σ x_i by default, or a Rastrigin function with cos(2π x_i) on [-5.12, 5.12]), and so does its f2 = g (1 - sqrt(f1 / g)) for CTP2–CTP7, so the reference values do not compare these problems.
  • Pareto-optimal sets: the paper places the Pareto-optimal solutions of CTP2–CTP5 on the line (f2 - e) cos(θ) = f1 sin(θ) (its eq. 6). They lie exactly on it only where the sine is 0: the 13 isolated points of CTP3 and CTP4, and 16 points of CTP5. CTP2's front is 13 short segments within 0.005 of the line. With c = 2, CTP5's constraint boundary leaves the line at f1 = 0 with a finite slope, so its first arch, up to f1 = 0.256, is Pareto-optimal as well. At such an isolated point, where d = 0.5, the square root turns the rounding of the sine, about 1e-14, into a G of about 1e-7: a design there is feasible only up to that tolerance.

VAMOS 1.2.0: only MW4, MW8 and MW14 accept an objective override; the other MW keys reject a different n_obj with ValueError. VAMOS 1.1.0 accepted an override for every MW key and kept those problems at two objectives, so the selection reported an objective count that the problem did not have.

VAMOS 1.2.0: cec2009_cf1 and vamos.problems.CEC2009CF1 default to n = 10, the dimension of the CF problems in the CEC2009 technical report and in PlatEMO; n_var still selects another one. The CEC2009_UF_CF_curved benchmark suite also runs CF1 with 10 variables. VAMOS 1.1.0 and earlier default to 30, so their default CF1 runs are not comparable with CEC2009 results. The reference values now compare CF1 with jMetal's CEC2009CF1 at commit 1e4b83cae6, which agrees with it. jMetal 7.8 has no CEC2009 CF classes (its CF1 is ZXH-CF1), so CF1 was not compared before.

In 1.0.2, cec2009_cf1 did not declare its constraint, so algorithms ran it as an unconstrained problem; it also stored the CEC2009 constraint c(x) >= 0 without changing its sign and searched [-1, 1] for variables 2 to n. VAMOS 1.1.0: CF1 declares its constraint, stores G = -c(x) so that G <= 0 is feasible, and uses the CEC2009 domain [0, 1]^n.

Binary, integer, mixed and permutation examples

These examples are useful for checking representations and operators. The synthetic instances have fixed generation seeds, independent of the optimizer's seed. Changing n constructs a different instance; it does not merely increase the search budget.

Key Encoding Default n / M G Default parameters Source Reference sample
zdt5 Binary 80 / 2 0 30-bit first block, ten 5-bit tail blocks ZDT5 2D, 31 points for default n
bin_feat Binary 50 / 2 0 Instance seed 12345 VAMOS synthetic None packaged
bin_knapsack Binary 50 / 2 0 Capacity ratio 0.4; instance seed 2023 VAMOS synthetic None packaged
bin_qubo Binary 30 / 2 0 Symmetric Q; instance seed 7 VAMOS synthetic None packaged
int_alloc Integer 20 / 2 0 Each allocation in 0..10; instance seed 321 VAMOS synthetic None packaged
int_jobs Integer 30 / 2 0 Five job labels 0..4; instance seed 99 VAMOS synthetic None packaged
mixed_design Mixed 9 / 2 0 Three real, three integer, three categorical; instance seed 123 VAMOS synthetic None packaged
tsp6 Permutation 6 / 2 0 Six equally spaced cities on the unit circle VAMOS TSP None packaged
kroa100, krob100, kroc100, krod100, kroe100 Permutation 100 / 2 0 Fixed packaged coordinates from the respective TSPLIB instance TSP formulation, coordinates None packaged

What the objectives actually mean

For binary x, each bit selects one item. Every expression below is minimized.

Key First objective Second objective Interpretation
bin_feat Negative sum of selected utilities Sum of selected feature costs Synthetic utility/cost trade-off; no classifier is trained
bin_knapsack Absolute difference between selected weight and target capacity Negative selected value Capacity deviation is an objective; overweight selections are allowed
bin_qubo xᵀQx + bᵀx Negative number of selected bits A two-objective synthetic QUBO extension
int_alloc Total allocation cost Negative sum of weighted square-root returns Diminishing returns; no global budget constraint
int_jobs Weighted mismatches against preferred job labels Largest job-type count divided by n Encourages a balanced use of labels; does not model machines or makespan
mixed_design Squared real/integer target errors plus category mismatches Integer and category costs Synthetic mixed-variable design trade-off
All TSP keys Sum of edge lengths in the closed tour Maximum edge length in that tour Total distance versus longest leg

TSP decisions must contain every city exactly once. Both objectives use the same coordinates and unrounded Euclidean distances, including the closing edge. A kroa100 run therefore uses TSPLIB coordinates in VAMOS's two-objective formulation; published single-objective TSPLIB tour costs with integer distance rounding are not directly comparable. The named tsp6 uses circle coordinates; constructing TSP() without arguments uses a separate asymmetric six-city toy.

ZDT5 requires n >= 35 and (n - 30) divisible by 5. Changing its number of tail blocks changes its front, so the packaged 80-bit reference is not universal. The TSPLIB factories have fixed dimensions. VAMOS 1.2.0: a different n_var is rejected with ValueError; VAMOS 1.1.0 ignored it while the selection reported it.

VAMOS 1.2.0: mixed_design rounds each categorical value and clips it to 0..cardinality - 1, as it does with integer values. VAMOS 1.1.0 wrapped an out-of-range category modulo its cardinality, so a value rounded to the cardinality counted as category 0. Values within the bounds evaluate as before.

Application-inspired benchmarks

RE: Tanabe–Ishibuchi engineering suite

These are fixed formulations: keep their dimensions as listed. G = 0 for all RE entries. Several include the sum of constraint violations as an additional objective; that trade-off differs from enforcing feasibility through G. The default bounds and discrete choices are fixed by the linked implementation.

Key / formulation Encoding n / M Violation objective Default parameters Source Reference sample
re21 — four-bar truss Real 4 / 2 No Fixed design bounds RE2 2D
re22 — reinforced concrete beam Mixed 3 / 2 f₂ 76 discrete reinforcement areas (see below) RE2 None packaged
re23 — pressure vessel Mixed 4 / 2 f₂ Thickness step 0.0625; two integer counts RE2 None packaged
re24 — hatch cover Real 2 / 2 f₂ x1 in [0.5, 4], x2 in [0.5, 50] RE2 2D
re25 — coil compression spring Mixed 3 / 2 f₂ Integer coil count; 19 wire diameters from 0.092 to 0.5 RE2 None packaged
re31 — two-bar truss Real 3 / 3 f₃ Fixed design bounds RE3a 3D
re32 — welded beam Real 4 / 3 f₃ Fixed design bounds RE3a 3D
re33 — disc brake Real 4 / 3 f₃ Fixed design bounds RE3a 3D
re34 — vehicle crashworthiness Real 5 / 3 No Fixed response surfaces RE3a 3D
re35 — speed reducer Mixed 7 / 3 f₃ One integer gear-teeth variable RE3b None packaged
re36 — gear train Integer 4 / 3 f₃ Integer tooth counts 12..60 RE3b None packaged
re37 — rocket injector Real 4 / 3 No Fixed response surfaces RE3b 3D
re41 — car side impact Real 7 / 4 f₄ Fixed response surfaces RE4 4D
re42 — conceptual marine design Real 6 / 4 f₄ Fixed design bounds RE4 4D
re61 — water resource planning Real 3 / 6 f₆ Fixed design bounds RE6/9 6D
re91 — car cab Real 7 / 9 No separate violation objective Deterministic mean-value evaluation (stochastic=False) RE6/9 9D

VAMOS 1.2.0: the RE and RWA entries, the TSPLIB instances and the application examples welded_beam, fs_real and ml_tuning reject an n_var other than the listed one with ValueError, as they already did for n_obj, and the RE registry labels name the formulations in this table. VAMOS 1.1.0 ignored such an n_var while the selection reported it: make_problem_selection("re21", n_var=10) reported 10 variables for a four-variable problem. Several RE labels named other formulations, and the re24 selection reported four variables for the two-variable hatch cover. Keep RE91's deterministic setting explicit when comparing it with stochastic formulations.

VAMOS 1.2.0: re24 and re25 use the domains of Tanabe and Ishibuchi's reference implementation (reproblem.py), which jMetal shares.

  • re24 searches x2 in [0.5, 50]. VAMOS 1.1.0 and earlier start at 4, which removes the low-cost designs that violate the shear-stress constraint: their front only reaches f₁ = 480.5, while the reference front starts at f₁ = 60.5.
  • re25 chooses among the 19 wire diameters from 0.092 to 0.5. The reference lists 42 diameters, but bounds the variable to [0.09, 0.5] and rounds it to the nearest listed value, so the 23 diameters below 0.092 are never used. VAMOS 1.1.0 and earlier allow all 42.

RE24 and RE25 results from those versions are not comparable with results on the reference problems.

re22 differs from the reference implementation in two documented ways:

  • Value list: its reinforcement areas read the reference's "3,10" as the single value 3.10, while reproblem.py and jMetal read two values, 3 and 10. The area 10 never reaches the front, and 3.10 adds a short segment to it, so the two fronts differ by at most 6e-4 after normalization.
  • Encoding: x1 is a categorical index into the 76 areas, while the reference rounds a continuous value in [0.2, 15] to the nearest area. Both reach the same areas, but variation operators sample them with different probabilities.

RWA: Zapotecas-Martínez and collaborators

All RWA entries use real variables, fixed dimensions and bounds, and zero separate G outputs. They evaluate application response functions. The implementation converts maximization goals to negative objective values.

Key / application n / M Default parameters Source Reference sample
rwa1 — honeycomb heat sink 5 / 2 Fixed response surface; maximize Nu, minimize friction factor RWA1–4 2D
rwa2 — vehicle crashworthiness 5 / 3 Fixed response surface RWA1–4 3D
rwa3 — synthesis gas production 3 / 3 Maximize CH₄ and CO; minimize H₂/CO RWA1–4 3D
rwa4 — wire electrical discharge machining 5 / 3 Maximize cutting rate; minimize roughness and dimensional deviation RWA1–4 3D
rwa5 — thermal storage 9 / 3 Fixed packed-bed response surface RWA5 3D
rwa6 — steel milling 4 / 3 z = 1, d = 1; maximize removal rate RWA6 3D
rwa7 — rocket injector 4 / 3 Three response objectives RWA7–8 3D
rwa8 — rocket injector 4 / 4 Four response objectives RWA7–8 4D
rwa9 — ultra-wideband antenna 10 / 5 Fixed passband/stopband response functions RWA9–10 5D
rwa10 — repellent fabric 3 / 7 Fixed fabric performance response functions RWA9–10 7D

rwa1, rwa6, rwa9 and rwa10 differ from jMetal's RWA implementations. VAMOS follows the formulas as published, but these four have not yet been checked against the article's appendix (doi 10.1016/j.engappai.2023.106192):

  • rwa1: the Nusselt-number coefficients use 4.148e-3; jMetal writes 4.148 * 10e-3 (4.148e-2), a likely transcription of "×10⁻³" that gives Nu around 1100–1900 instead of the physically plausible 79–305.
  • rwa6: the removal rate uses d = 1, jMetal d = 2.5. The constant only scales f3, so dominance is the same, but values and fronts are not comparable with jMetal's.
  • rwa9: the ten geometric variables of the antenna (d1 enters no formula); jMetal has six variables and derives a2 = l1 * w1 * l2 * w2.
  • rwa10: no additive constants beyond the response functions; jMetal adds constants to several objectives (−15.33, −29.33, −413.33, −73.33, +0.22, −8413.33, −2814.83). That translates the objectives without changing dominance, but it changes absolute values and reference points.

VAMOS 1.2.0: rwa7 and rwa8 search [0, 1]^4, the normalized domain in which Goel et al. (2007) fitted the injector response surfaces, as re37, Tanabe and Ishibuchi's RE37 and jMetal's RWA7 and RWA8 do. The last TTmax term is -0.281 * dHA * alpha * dOA, as in re37, so rwa7 is re37 with its objectives in the order Xcc, TFmax, TTmax. VAMOS 1.1.0 and earlier evaluate the polynomials on the physical design ranges (alpha in [0, 20], dHA in [0, 25], dOA in [-40, 0], OPTT in [0.01, 0.02]), far outside the fitted region, with the opposite sign in that term: TTmax ranges from about -2242 to 17093 instead of -0.43 to 1.10. RWA7 and RWA8 values, fronts and indicators from those versions are not comparable with the published problems.

The packaged rwa7.csv (3000 points) and rwa8.csv (3999 points) samples are regenerated for the normalized formulation. They come from dense sampling of [0, 1]^4 with local refinement, each selected design moved to the front by minimizing an achievement function, and a non-dominated filter over all the evaluated designs. No point of either sample is dominated by jMetal's RWA7 and RWA8 reference fronts or by Tanabe and Ishibuchi's RE37 reference points.

CRE and RCM: constrained real-world suites

These problems keep their constraints in G (G <= 0 is feasible) instead of adding the violation as an objective, as RE does. They follow jMetal's classes at commit 1e4b83cae6 (CRE, RCM), which the reference values check. All are real-valued, with fixed n and M. The integer variables marked rounded stay real; the evaluation rounds them, as jMetal does.

CRE are Tanabe and Ishibuchi's constrained problems (2020):

Key / formulation n / M / G Notes Reference sample
cre21 — two-bar truss 3 / 2 / 3 RE31 with its constraints; same as rcm03 2D
cre22 — welded beam 4 / 2 / 4 RE32 with its constraints 2D
cre23 — disc brake 4 / 2 / 4 RE33 with its constraints; same as rcm05 2D
cre24 — speed reducer 7 / 2 / 11 RE35 with its constraints; x3 rounded, halves to even 2D
cre25 — gear train 4 / 2 / 1 RE36 with its constraint; all rounded, halves to even 2D
cre31 — car side impact 7 / 3 / 10 Same as car_side_impact 3D
cre32 — conceptual marine design 6 / 3 / 9 RE42 without RE42's guards 3D
cre51 — water resource planning 3 / 5 / 7 Same as water and rcm11 5D

RCM are the 50 problems of the CEC 2021 real-world constrained suite (Kumar et al., 2021):

Key / formulation n / M / G Notes Reference sample
rcm01 — pressure vessel 4 / 2 / 2 x1 and x2 rounded 2D
rcm02 — vibrating platform 5 / 2 / 5 2D
rcm03 — two-bar truss 3 / 2 / 3 Same as cre21 2D
rcm04 — welded beam 4 / 2 / 4 cre22 with another buckling load 2D
rcm05 — disc brake 4 / 2 / 4 Same as cre23 2D
rcm06 — speed reducer 7 / 2 / 11 x3 rounded 2D
rcm07 — gear train 4 / 2 / 1 Not rounded 2D
rcm08 — car side impact 7 / 3 / 9 3D
rcm09 — four-bar plane truss 4 / 2 / 0 No constraints 2D
rcm10 — two-bar plane truss 2 / 2 / 2 2D
rcm11 — water resource management 3 / 5 / 7 Same as cre51 5D
rcm12 — simply supported I-beam 4 / 2 / 1 2D
rcm13 — gear box 7 / 3 / 11 x3 rounded 3D
rcm14 — multiple-disk clutch brake 5 / 2 / 8 2D
rcm15 — spring 3 / 2 / 8 x1 and x3 rounded; x3 picks one of 42 wire diameters 2D
rcm16 — cantilever beam 2 / 2 / 2 2D
rcm17 — bulk carrier 6 / 3 / 9 f1 is unbounded below on the feasible set None packaged
rcm18 — front rail 3 / 2 / 3 2D
rcm19 — multi-product batch plant 10 / 3 / 10 x1–x3 rounded 3D
rcm20 — hydrostatic thrust bearing 4 / 2 / 7 2D
rcm21 — crash energy management 6 / 2 / 4 2D
rcm22 — Haverly's pooling 9 / 2 / 6 Four equalities 2D
rcm23 — reactor network 6 / 2 / 5 Four equalities None packaged
rcm24 — heat exchanger network 9 / 3 / 6 Six equalities 3D
rcm25 — process synthesis 2 / 2 / 2 x2 rounded 2D
rcm26 — process synthesis and design 3 / 2 / 2 x3 rounded; one equality None packaged
rcm27 — process flow sheeting 3 / 2 / 3 2D
rcm28 — two-reactor problem 7 / 2 / 8 x5 and x6 rounded; four equalities 2D
rcm29 — process synthesis 7 / 2 / 9 x4–x7 rounded 2D
rcm30–rcm32 — SOPWM of 3-, 5- and 7-level inverters 25 / 2 / 24 Angles in degrees, increasing 2D, each
rcm33–rcm35 — SOPWM of 9-, 11- and 13-level inverters 30 / 2 / 29 Angles in degrees, increasing 2D, each
rcm36–rcm38 — distributed generation sizing 28 / 2 / 24 24 equalities 2D, each
rcm39 — distributed generation sizing 28 / 3 / 24 24 equalities 3D
rcm40, rcm42, rcm43 — optimal power flow 34 / 2 / 26 26 equalities 2D, each
rcm41, rcm44, rcm45 — optimal power flow 34 / 3 / 26 26 equalities 3D, each
rcm46 — optimal power flow 34 / 4 / 26 26 equalities 4D
rcm47, rcm48 — islanded microgrid droop control 18 / 2 / 12 12 equalities 2D, each
rcm49 — islanded microgrid droop control 18 / 3 / 12 12 equalities 3D
rcm50 — power distribution system planning 6 / 2 / 1 One equality None packaged

VAMOS 1.2.0: CRE21–CRE51 and RCM01–RCM50 follow jMetal's classes at commit 1e4b83cae6. That includes jMetal's pull requests #513 (CRE21, CRE25 and CRE31 as in the authors' reference code) and #514 (RCM10's load P = 104, and the inequalities of RCM26 and RCM28 before their equalities).

  • Overlap: CRE21, CRE23 and CRE51 are RCM03, RCM05 and RCM11: the same bounds, objectives and constraints. CRE31 is car_side_impact and CRE51 water. The other pairs differ:
  • CRE22 and RCM04 in the buckling load: x3^2 x4^6, against the suite's x3^2 + x4^6.
  • CRE24 and RCM06 in the bounds and rounding of x3, and in the bound on the stress of the second shaft (1100 against 850).
  • CRE25 and RCM07 in the bounds and the rounding of the numbers of teeth.
  • CRE31 and RCM08 in f1 and one constraint, as the suite's code writes them; RCM08 also lacks CRE31's bound on the velocity of the front door.
  • CRE32 and RCM17 in their models of the bulk carrier.
  • Equalities: as in jMetal, each equality h(x) = 0 of RCM22–RCM24, RCM26, RCM28 and RCM36–RCM50 is the constraint |h(x)| - 1e-4 <= 0. Uniform random designs never meet them.
  • Rounded variables: the integer and binary variables stay real, and the evaluation rounds them as jMetal does:
  • CRE24 and CRE25 with Math.rint, halves to even;
  • RCM01, RCM06, RCM13, RCM15, RCM19, RCM25, RCM26, RCM28 and RCM29 with Math.round, halves up.

RCM07 and RCM27 are not rounded, in jMetal or in the suite's code, although their bounds of ±0.49 around integers suggest it. - RCM09 has no constraints: the suite's metadata lists one, but its evaluator never sets it. - Singular points: RCM20 divides by zero where r = ro, so f2 is infinite there, as in jMetal. VAMOS rejects that evaluation as non-finite. Random designs and five NSGA-II runs of 30 000 evaluations did not reach it. - Fronts: RCM17 has none: its f1 is unbounded below on the feasible set. Feasible designs dominate rows of jMetal's fronts of RCM23, RCM26 and RCM50, so those are not packaged (see Reference fronts).

Learning and mixed engineering examples

Install python -m pip install "vamos-optimization[examples]==1.4.0" for the two classifier examples. The base package is sufficient for welded_beam. See Installation for environments and extras.

Key Encoding Default n / M G Defaults and objective meaning Source Reference sample
fs_real Binary 30 / 2 0 Breast-cancer data; stratified 70/30 split, random_state=0; logistic-regression validation error vs selected-feature count Feature selection None packaged
ml_tuning Mixed 4 / 2 0 Same dataset/split; SVC validation error vs support-vector count plus 0.1 × degree SVM tuning None packaged
welded_beam Mixed 6 / 2 4 Four real dimensions, one material category, one stiffener count; fabrication cost vs deflection Engineering example None packaged

fs_real penalizes the empty mask with error 1 and feature-count objective n. ml_tuning searches C in [10⁻², 10²], gamma in [10⁻³, 10¹], kernels rbf/linear/poly and degrees 2–5. These are validation objectives, not unbiased held-out test estimates. Both public classes also accept the wine dataset, which changes fs_real to 13 decision variables; named registry selection uses the defaults above.

VAMOS 1.2.0: when the classifier of fs_real or ml_tuning fails to train on a candidate with a data error (ValueError or LinAlgError), that candidate gets validation error 1 and the evaluation warns with the number of failed models. If every model of an evaluation fails, it raises EvaluationError, and any other exception propagates. VAMOS 1.1.0 turned every exception into error 1 without warning, so a bug in the model looked like a poor candidate.

VAMOS 1.1.0 constraint metadata: WeldedBeamDesignProblem declares its four constraint columns, so registry-created problems enforce them. In 1.0.2, an explicit problem.n_constraints = 4 was required; that assignment remains harmless. Follow the engineering lesson and inspect the returned G before selecting a design.

welded_beam is VAMOS's six-variable mixed design example; it is a different formulation from the four-variable, three-objective re32. Its four emitted G columns are shear stress, normal stress, deflection and the h <= b condition. Its formulas are not those of the four-variable welded-beam benchmark, which re32 and cre22 use. With the same h, l, t and b:

  • the weld term of the cost is 1.10471 h^2, without the factor l; the cost is then scaled by the material factor (1, 1.15 or 1.3) and adds 5 per stiffener;
  • the deflection 2.1952 / (t^3 b) is also divided by the stiffener count, in the objective and in its constraint;
  • the normal stress is 504000 / (t b^2) instead of 504000 / (t^2 b);
  • the polar moment is J = 2 sqrt(2) h l (l^2/4 + R^2) instead of 2 sqrt(2) h l (l^2/12 + ((h + t)/2)^2), and the shear stress combines the primary and torsional stresses with the cross term 2 tau' tau'' instead of tau' tau'' l / R;
  • there is no buckling constraint.

Its results are therefore not comparable with published welded-beam results. See Known limitations.

VAMOS 1.2.0: cec2009_uf3 searches the CEC2009 domain [0, 1]^n, as in the CEC2009 technical report, jMetal and the MOEA Framework. VAMOS 1.1.0 and earlier search [-1, 1] for variables 2 to n. The Pareto front is the same, but the larger search space changes the problem's difficulty.

VAMOS 1.2.0: cec2009_uf9 adds the squared distance y_j^2 of each remaining variable, as in the CEC2009 technical report, jMetal and PlatEMO, and as cec2009_uf8 does. VAMOS 1.1.0 and earlier use the multimodal distance 2 y_j^2 - cos(4 pi y_j) + 1 of UF5. The Pareto front is the same, since y_j = 0 there, but the landscape differs, so UF9 results from those versions are not comparable with CEC2009 results.

Reference fronts

The packaged CSV directory contains finite reference samples. Check the CSV's number of columns, objective definitions, signs, scaling and parameter settings against the run. A file's existence alone is insufficient evidence of compatibility or a certified optimum.

  • ZDT fronts have two columns; ZDT5's sample is tied to its default 80 bits.
  • DTLZ fronts have three columns, and dtlz<k>.2d.csv two. Other values of M need a supplied reference. Development after 1.4.0: DTLZ1–DTLZ4 and DTLZ7 also have samples with 4, 6 and 8 objectives, dtlz1.<M>d.csv, dtlz2.<M>d.csv and dtlz7.<M>d.csv. DTLZ3 and DTLZ4 share DTLZ2's front, so they resolve to dtlz2.<M>d.csv. DTLZ5 and DTLZ6 with more than three objectives, and the other values of M, still need a supplied reference.
  • WFG and LSMOP samples have two columns, even though their registered default is three objectives. Supply a matching reference for default runs; do not use those two-column files for three-objective indicators.
  • Development after 1.4.0: CTP1–CTP7 have two-objective samples, ctp1.csv to ctp7.csv.
  • Development after 1.4.0: IDTLZ1 and IDTLZ2 have samples with two and three objectives, idtlz1.2d.csv, idtlz1.3d.csv, idtlz2.2d.csv and idtlz2.3d.csv. Other values of M need a supplied reference.
  • C-DTLZ, DC-DTLZ, ZCAT and ZXH-CF supply dimension-specific samples, named <key>.<M>d.csv: at M = 2 and 3 for C-DTLZ and DC-DTLZ, at M = 2, 3, 4 and 6 for ZCAT, and at M = 3 for ZXH-CF. Other objective counts need a supplied reference.

VAMOS 1.2.0: the automatic lookup takes <key>.<M>d.csv or, failing that, <key>.csv, for the run's number of objectives or, when none is given, the problem's registered default. It never returns a file with another number of objectives: WFG and LSMOP at their default three objectives, DTLZ at two, and C-DTLZ and DC-DTLZ at three, have no automatic reference. VAMOS 1.1.0 returned the file it found, whatever its columns, so its automatic references for those cases had the wrong number of objectives. The C-DTLZ and DC-DTLZ files of 1.1.0 (c1dtlz1.csv and so on, two-objective) are replaced by <key>.2d.csv and <key>.3d.csv.

VAMOS 1.3.0: DTLZ1–DTLZ7 also have two-objective samples, dtlz1.2d.csv to dtlz7.2d.csv, so a two-objective DTLZ run has an automatic reference. VAMOS 1.2.0 had none, and VAMOS 1.1.0 returned the three-objective file. - RE/RWA samples have the dimensions shown above. Their presence does not imply that every sample point is a certified global Pareto optimum. - Families marked “None packaged” require a suitable external reference or a documented empirical reference construction when using reference-front metrics.

VAMOS 1.2.0: every packaged sample is mutually non-dominated, and the analytic ones lie on their fronts within 1e-9. The table gives the origin of each file. experiments/scripts/generate_reference_fronts.py regenerates the files marked as regenerated.

Files Origin
ZDT1.csv, zdt2–zdt4, zdt6 Regenerated: 1000 points spread along the closed-form front
ZDT5.csv 31 exact points for the default 80 bits, unchanged
dtlz1–dtlz4 Regenerated: 5050-point simplex lattice, scaled to sum(f) = 0.5 (DTLZ1) or projected onto the unit sphere
dtlz5, dtlz6 Regenerated: 1000 points of the front curve
dtlz7 Regenerated: 5050 points of the disconnected front
dtlz1.2d–dtlz7.2d VAMOS 1.3.0: regenerated, 1000 points spaced evenly along the front: f1 + f2 = 0.5 (DTLZ1), the quarter circle (DTLZ2–DTLZ6; at M = 2, DTLZ5 and DTLZ6 do not degenerate) and the non-dominated pieces of f2 = 4 - f1 (1 + sin(3 pi f1)) (DTLZ7)
idtlz1, idtlz2, .2d and .3d Development after 1.4.0: generated as PlatEMO's GetOptimum defines them, (1 - w) / 2 (IDTLZ1) and 1 - w / norm(w) (IDTLZ2): at M = 2, 1000 points spaced evenly along the front (IDTLZ1's segment is DTLZ1's); at M = 3, the 5050-point simplex lattice of the DTLZ1–DTLZ4 samples
dtlz1, dtlz2 and dtlz7, .4d, .6d and .8d Development after 1.4.0: regenerated. DTLZ1 and DTLZ2 (which DTLZ3 and DTLZ4 share): the largest Das-Dennis simplex lattice with at most 10 000 points (H = 37, 13 and 8: 9880, 8568 and 6435 points), scaled to sum(f) = 0.5 or projected onto the unit sphere. DTLZ7: its last objective is 2 (M - sum psi(f_i)) over the others, with psi(t) = t (1 + sin(3 pi t)), so the front is the product of the values of f_i that no smaller value beats in psi; values of a grid of 200 001 points, combined and thinned to 5050 points by farthest-point selection
wfg1, wfg2 Regenerated from the WFG shape functions, 10 000 points
wfg3–wfg9 Closed-form front, 10 000 points, unchanged
lsmop1–lsmop8 Closed-form front, 1000 points, unchanged
lsmop9 Regenerated: 1000 non-dominated points of the closed-form front
cec2009_uf1–uf5, uf7, uf8, uf10 Closed-form front, unchanged
cec2009_uf6 Closed-form front with its isolated point (0, 1) added
cec2009_uf9 Regenerated: 5111 points of UF9's two planar pieces, sum(f) = 1 with f1 <= (1 - f3)/4 or f1 >= 3(1 - f3)/4
mw1–mw3, mw5–mw7 Unchanged; within a mean distance of 0.0085 of pymoo 0.6.2's fronts
binh2, golinski, osyczka2, water, car_side_impact jMetal's fronts (commit 1e4b83cae6), under LICENSE-jmetal.txt
srinivas, constrex, tanaka Regenerated: the feasible, non-dominated designs of the Pareto-set pieces and the constraint boundaries that hold the front, as VAMOS evaluates them, 1000 points. jMetal's samples have dominated points: 310 of Srinivas's 3794 are dominated by another of its points or by a random feasible design, and 774 by those or by the new sample
mw4, mw8, mw14 (.2d and .3d), mw9–mw13 jMetal's fronts (resources/referenceFronts at commit 1e4b83cae6), under LICENSE-jmetal.txt; jMetal's MW problems are VAMOS's, which the reference values check
cec2009_cf2–cec2009_cf10 jMetal's fronts (resources/referenceFronts at commit 1e4b83cae6), under LICENSE-jmetal.txt; a feasible design on the Pareto set reaches every point. CF4 and CF5, CF6 and CF7, and CF9 and CF10 share their fronts
zxh_cf1.3d–zxh_cf16.3d Regenerated: on the Pareto set S = 1 and h = 0, so the objectives are the front shape of the angles; designs on a 301 × 301 grid of angles, evaluated by VAMOS's problem, keep the feasible, non-dominated ones, thinned to 5050 points. A feasible design reaches every point within 1e-9 (see the notes on jMetal's fronts)
ctp1–ctp7 Development after 1.4.0: generated from the paper's definition at g = 1 and on the constraint boundaries, as VAMOS evaluates them. CTP1: the envelope max(exp(-f1), a_j exp(-b_j f1)), 1000 points. CTP2–CTP7: the feasible, non-dominated points of the boundary, sampled at 2 000 001 values of sin(θ) (f2 - e) + cos(θ) f1 and at the zeros of its sine, of the unconstrained front f2 = 1 - f1 and of the least feasible f2 at f1 = 0 and 1, thinned to at most 1000 with the zeros kept: CTP2's 13 segments, the 13 isolated points of CTP3 and CTP4 (the same file), CTP5's first arch and 15 points, CTP6's boundary in 1 <= sin(θ) (f2 - e) + cos(θ) f1 <= 2, and six pieces of CTP7's unconstrained front with (0, 1.0446), the least feasible point at f1 = 0. A feasible design reaches every point (within 1e-6 at the isolated points; see the notes on CTP above)
lircmop1–lircmop14, dascmop1–dascmop9 jMetal's fronts (resources/referenceFronts at commit 1e4b83cae6), under its MIT notice in LICENSE-jmetal.txt; DAS-CMOP8 and DAS-CMOP9 leave out two rows each, which another row dominates
cre21–cre25, cre31, cre32, cre51, rcm01–rcm49 except rcm17, rcm23 and rcm26 jMetal's fronts (resources/referenceFronts at commit 1e4b83cae6), unchanged, under LICENSE-jmetal.txt; approximations (see the notes on jMetal's CRE and RCM fronts). cre51 and rcm11 are the same file
c1dtlz1–dc3dtlz3, .2d and .3d, c3dtlz4 included Regenerated: the feasible, non-dominated part of DTLZ's front (C3-DTLZ1 and C3-DTLZ4: of their constraint boundaries), as VAMOS evaluates it; 1000 points at M = 2 and up to 5050 at M = 3
zcat* Unchanged samples
jmetal-zdt1–jmetal-zdt4, jmetal-zdt6, jmetal-dtlz1, jmetal-dtlz2 and jmetal-dtlz5–jmetal-dtlz7, .2d and .3d VAMOS 1.3.0: jMetal's fronts (resources/referenceFronts at commit 1e4b83cae6), unchanged, under LICENSE-jmetal.txt; the optional jmetal source described below. jMetal's DTLZ3 and DTLZ4 files are its DTLZ2 file, so jmetal-dtlz2 serves the three
re21, re24, re31–re34, re37, re41, re42, re61 Tanabe and Ishibuchi's approximated fronts (reference_points_RE*.dat)
re91 Regenerated by dense sampling of the deterministic RE91, 1806 points
rwa1, rwa2, rwa4–rwa6 Unchanged samples
rwa3, rwa9, rwa10 Regenerated by dense sampling, 1484, 1343 and 1493 points
rwa7, rwa8 Regenerated for the normalized formulation (see the RWA notes above)

VAMOS 1.3.0: jMetal's fronts of ZDT1–ZDT4, ZDT6 and DTLZ1–DTLZ7 at M = 2 and 3, the files Evolver also uses, are packaged as an optional source. VAMOS's own samples stay the default. Both kinds agree on the hypervolume (within 0.02 % for ZDT and 1.6 % for DTLZ, measured at the nadir plus 10 % of the range of the two together), but their points lie at different places along the front, so IGD, IGD+ and Δp computed against one differ from those computed against the other. Choose jMetal's to compare those indicators with jMetal's and Evolver's:

  • The files are jmetal-<key>.csv (ZDT) and jmetal-<key>.<m>d.csv (DTLZ; jmetal-dtlz2.<m>d.csv for DTLZ2, DTLZ3 and DTLZ4, whose jMetal files are the same). For a problem or objective count that jMetal's set lacks there is no jMetal front, and VAMOS never falls back to its own file.
  • vamos --hv-threshold 0.9 --hv-reference-front jmetal measures the stopping target against jMetal's front; vamos names VAMOS's front and any other value is a CSV path. The run records reference_front_source (vamos, jmetal or file) next to reference_front_path in its resolved termination.

Development after 1.4.0: the lattices of DTLZ1–DTLZ4 with 4, 6 and 8 objectives are those of PlatEMO's reference fronts (UniformPoint(10000, M), which also raises zeros to 1e-6). jMetal's DTLZ2, DTLZ3 and DTLZ4 files with those numbers of objectives (DTLZ2.4D.csv and so on) hold the same points, in the same order, so the default source gives the indicator values of jMetal and Evolver for these problems. The jmetal source keeps two and three objectives: jMetal's DTLZ1 and DTLZ7 files with 4, 6 and 8 objectives are sparser (183 to 370 points; DTLZ1's are rounded to five decimals and stay away from the front's corners), and VAMOS does not package them.

Notes on the table:

  • Dense sampling: for RE91, RWA3, RWA9 and RWA10, the generator
  • evaluates a dense sample of the design box, refines locally around the non-dominated designs, and filters every evaluated design for non-dominance;
  • moves a well-spread pool of those designs to the front by minimizing an augmented achievement function;
  • keeps the polished designs that stay non-dominated, and spreads the packaged points by farthest-point selection.
  • jMetal's ZXH-CF fronts: only CF7's lies on the feasible front. The others approximate it, so VAMOS generates all sixteen. Their mean distance to VAMOS's samples is 3e-3 to 7e-3. On the Pareto set S = 1 and h = 0, so each front point gives its angles and a design on the Pareto set at those angles. Against those designs:
  • CF2, CF8, CF11 and CF13 (concave) lie inside the unit sphere, by 2e-5 to 5.5e-4, where no design reaches.
  • CF3, CF5, CF10 and CF16 (convex) lie behind the front, by up to 1.9e-3, so the designs dominate them.
  • CF4, CF6, CF9 and CF15 (mixed) are off in f1 by up to 1.2e-2, and the designs dominate 709 to 852 of their 1000 points.
  • CF1, CF12 and CF14 (linear) lie behind the front by up to 1.2e-6.
  • CF9, CF10, CF11, CF12 and CF14 have one to three points at excluded angles, which no feasible design reaches.

200 000 random designs already dominate points of CF3, CF4, CF6 and CF16. - Tanabe and Ishibuchi's fronts: they are redistributed under the MIT license of their repository, whose notice ships next to them as LICENSE-reproblems.txt. Their values are unchanged. They approximate the fronts, so a dense random sample of a problem still dominates a small share of their points: about 6 % for RE61, 2.5 % for RE33, and at most 0.5 % for the others. RE91's front from the same source does not apply: it comes from the stochastic formulation, while VAMOS evaluates RE91 at the mean of its stochastic parameters. - jMetal's CRE and RCM fronts: they approximate the fronts: the non-dominated feasible solutions of many runs of several algorithms, polished near their extremes and, where the equalities leave thin feasible sets, by problem-specific searches. No file has a duplicate row or a row that another dominates, and 200 000 uniform random designs dominate none of their rows. Those designs are seldom feasible, though: none is for CRE21, RCM03, RCM22, RCM24 and RCM28–RCM49. Random designs on the feasible sets of CRE21 and RCM03 (small bar areas), RCM22 (on its equalities), RCM28 (one reactor), RCM29 (small flows) and RCM30–RCM35 (increasing angles) dominate none either, beyond the ten significant digits of the files. Such designs do dominate rows of three of jMetal's fronts, which are therefore not packaged: - RCM23: designs on its equalities improve 28 of its 94 rows, by up to 3.5 % of the front's range in both objectives; - RCM50: designs on its power balance improve one of its 104 rows by 0.7 % of the range; - RCM26: designs on its equality improve 163 of its 1000 rows, by up to 4.4e-5 of the range.

CRE23 and RCM05 are the same problem with different fronts; each front dominates 35 to 42 rows of the other, by up to 2e-4 of the range. RCM47's front is a single point. - Replaced samples: VAMOS 1.1.0 and earlier packaged inaccurate samples for several regenerated files, so IGD, IGD+, epsilon and normalized HV values computed with those files are not comparable across versions: - DTLZ and ZDT points up to 0.108 and 0.015 off their fronts; - a WFG1 front shifted by +0.0695 in both objectives; - DC2-DTLZ fronts on an infeasible layer at g ≈ 62; - two-objective C-DTLZ and DC-DTLZ files that the automatic lookup returned for the problems' default three objectives; - UF9's file was a copy of UF8's sphere; - LSMOP9 and the sampled RE and RWA files had dominated points.

The CLI accepts --hv-reference-front path/to/front.csv with --hv-threshold; use one column per objective and validate the selected file before launching a study. A reference front is a set of objective vectors; a hypervolume reference point is one vector bounding the evaluated region. See Analysis for indicators and comparison guidance.

CLI presets

Preset Members
families zdt1, dtlz2, wfg4, tsp6 (includes different encodings)
zdt ZDT1/2/3/4/6; binary ZDT5 is selected separately
dtlz, wfg, zcat, lz, lsmop The corresponding suites above
cec CEC2009 UF1–10 and CF1
constrained_many The 10 C-/DC-DTLZ and 14 MW cases
tsp tsp6
tsplib kroa100, krob100, kroc100, krod100, kroe100
real_world ml_tuning, welded_beam, fs_real; this preset does not include RE/RWA

Use --problem-set <preset>. A shared dimension override must make sense for every member; prefer explicit problem selections for studies that mix fixed and scalable formulations. For a new objective function, follow Build your own problem.

For verified application walkthroughs, see feature selection, model hyperparameters, and decision representations.