Bergstra et al. 2011: sample-efficient sequential optimizer that's the
workhorse of Hyperopt and Optuna. Different surrogate from BO's
Gaussian process — TPE models p(x | y < y*) with one KDE and
p(x | y >= y*) with another, then samples candidates from the 'good'
KDE and ranks by the ratio l(x) / g(x). The acquisition is implicit
in the ratio (a closed-form analog of Expected Improvement).
Implementation:
- 1-D Gaussian KDE per axis, with bandwidth chosen by Scott's rule
- Per-step:
- Evaluate observations into 'good' (top γ fraction by target) and
'bad'
- Sample n_candidates from the good distribution (independent per
axis) and pick the one with the largest l(x)/g(x)
- Evaluate it, append to history
Vec<f64> only, single-objective only. Compared with BayesianOpt:
- Cheaper per-step (no GP factorization)
- Doesn't need kernel hyperparameter tuning to work well
- Naturally extends to mixed/categorical decision types (future work)
- Generally less sample-efficient than well-tuned BO on smooth
continuous problems, but more robust out of the box
Tests cover convergence on 1-D Sphere within a tight budget,
deterministic reruns, panic on multi-objective.
Wierstra et al. 2008/2014 NES with the diagonal-covariance "separable"
variant (sNES). Different theoretical foundation from CMA-ES: rather
than tracking a full covariance matrix and adapting it through
evolution paths, sNES updates the sampling distribution's parameters
by following the natural gradient of expected fitness.
Each generation:
- Sample λ offspring from N(μ, diag(σ²))
- Rank-shape the fitnesses (utility weights from the standard NES table)
- Update μ along the natural gradient: μ ← μ + η_μ · σ · sum(u_i · z_i)
- Update σ multiplicatively: σ_j ← σ_j · exp(η_σ/2 · sum(u_i · (z_i,j² - 1)))
Vec<f64> decisions only, single-objective only. The diagonal covariance
makes per-step cost O(λ·n) instead of CMA-ES's O(λ·n²) — much faster on
high-dimensional problems where full-covariance tracking is expensive
or numerically fragile, at the cost of being unable to handle strongly
rotated landscapes.
Adds runners for the four expensive-eval / gradient-free additions to
the appropriate single-objective sections of `examples/compare.rs`:
- Rastrigin (multimodal): now also shows IPOP-CMA-ES alongside vanilla
CMA-ES so the restart benefit is directly visible.
- Rosenbrock (smooth valley): adds Nelder-Mead (well-suited) and (1+1)
ES (cheap baseline).
- Ackley + Rosenbrock: BayesianOpt run with a deliberately TINY budget
(60 evaluations vs 30k for the population-based methods) so the
sample-efficiency claim is visible — BO with 60 evals vs DE/CMA-ES
with 30k.
The compare harness now sides-by-sides 23 algorithms total across the
seven benchmark problems.
The first sample-efficient algorithm in heuropt. Bayesian optimization
maintains a Gaussian-process surrogate of the objective and at each
step picks the next decision by maximizing an acquisition function on
that surrogate, so the evaluation budget is used surgically.
Implementation:
- **Kernel**: anisotropic RBF (squared-exponential) with per-axis
length scales, signal variance, and a small noise/jitter floor.
Hyperparameters are exposed in the config; a future version can add
marginal-likelihood maximization.
- **Posterior**: standard formulation. Cholesky factorizes K (using
the new internal helper); mean and variance predictions follow.
- **Acquisition**: Expected Improvement against the best observed
feasible point. Optimized by best-of-N random sampling — simple,
predictable cost, no inner-optimizer footgun.
- **Initial design**: `initial_samples` uniform-random points in
bounds before the BO loop starts.
- **Constraints**: feasibility-aware EI — best observed value uses
only feasible points; infeasible candidates are penalized.
Vec<f64> decisions, single-objective only. Targets the regime no
existing heuropt algorithm covers: 50–500 evaluations on an
expensive black-box function (CFD sim, ML training run, real-world
measurement).
Tests cover convergence on the 1-D sphere within a tight evaluation
budget (~30 evals get to f < 1e-6 — vs population-based methods
needing thousands), deterministic reruns, and panic on
multi-objective + dim mismatches.
Hand-rolled `A = L · L^T` factorization plus forward/backward triangular
solves, used by the upcoming Bayesian Optimization implementation for
the GP posterior. Same f64 row-major Vec<Vec<f64>> interface as the
existing Jacobi eigen helper so we don't pull in nalgebra for one
algorithm.
Returns Err on non-positive-definite input (a small jitter is the
typical caller-side fix). Tested against the standard 2x2 case, the
3x3 known-result case, A·x = b round-trip, and the SPD-failure case.
Auger & Hansen 2005 IPOP-CMA-ES: wraps the existing CmaEs in a restart
loop that doubles the population size and re-randomizes the mean
whenever a restart trigger fires. Specifically addresses the failure
mode we observed on Rastrigin (vanilla CMA-ES = 2.3 vs DE = 0).
Restart triggers:
- The whole budget for one inner CmaEs run finishes without improvement
- (More sophisticated triggers — eigenvalue collapse, condition-number
blow-up, sigma stagnation — are left for future versions; the
per-run budget trigger captures the bulk of the practical benefit)
Each restart:
- Doubles the population_size (Auger & Hansen 2005)
- Re-randomizes the initial mean to a fresh point in the bounds box
- Resets sigma to the user's initial value
Same Vec<f64> + single-objective constraints as CmaEs. The total
budget is divided across restarts; restart budget grows with
population. Tests verify it beats vanilla CMA-ES on Rastrigin.
Nelder & Mead 1965: gradient-free local optimizer that maintains a
simplex of n+1 points in n-D and at each iteration replaces the worst
vertex by one of {reflect, expand, outside-contract, inside-contract,
shrink} relative to the centroid of the rest. The five standard
coefficients (reflection α=1, expansion γ=2, contraction ρ=0.5,
shrinkage σ=0.5) are exposed in the config but default to canonical
values so users can leave them alone.
Single-objective only, Vec<f64> only, bounds enforced by clamping
each new vertex. Termination is purely iteration-count for v0.2;
"vertices have collapsed" stopping is a future enhancement.
Filling a real gap: heuropt had population-based local search
(SimulatedAnnealing, HillClimber) but no classical direct-search
algorithm. Excellent for low-dim smooth-ish problems where a
population is overkill.
Rechenberg 1973's elemental evolution strategy: one parent, one child
each generation, accept the child if it is no worse, and adapt the
mutation step size by tracking the success rate. If more than 1/5 of
recent moves were accepted the search is too cautious — multiply σ by
`step_increase` (typical 1.22). Below 1/5 — divide by the same factor.
At 1/5 — leave it alone. The success window has length `adaptation_period`.
Single-objective only. Vec<f64> only. Generic Gaussian step bounded by
the embedded `RealBounds`.
Why ship it: it's the smallest possible self-adapting evolution strategy
and a useful pedagogical / baseline endpoint. Pairs well as the budget
floor ("give me anything cheaper than CMA-ES").
Expands the comparison harness with four new test problems chosen for
their distinct geometry:
- **Rosenbrock** (single-obj, smooth valley): the classic non-convex
smooth function. Differentiates CMA-ES (which exploits the local
metric) from Rastrigin's multimodal-trap regime.
- **Ackley** (single-obj, exponential multimodal trap): a more
forgiving multimodal test than Rastrigin — fewer narrow local
minima — so CMA-ES can show its strength while DE/GA still win.
- **ZDT3** (multi-obj, disconnected front): the only ZDT-family
problem with a non-contiguous Pareto front. Tests an algorithm's
ability to maintain spread across gaps.
- **DTLZ1** (many-obj, 3-D linear front): a triangular plane in
objective space (vs DTLZ2's spherical octant). Different shape
reveals which many-obj algorithms are biased toward sphere-like
fronts vs which infer geometry adaptively.
Each new section runs all applicable algorithms × N seeds × the
algorithm-class budget the existing sections already use.
Zhang, Tian & Jin 2015 KnEA: many-objective MOEA that biases survival
selection toward 'knee points' on the Pareto front — points where a
small improvement in one objective costs a large degradation in
another.
Each generation:
- NSGA-II-like loop with offspring + non_dominated_sort
- For the splitting front, identify knee points by perpendicular
distance from the hyperplane connecting the front's extreme points.
Members further from the hyperplane (= more 'kneeness') are preferred.
- Survival keeps every knee-tagged member; if room remains, fill from
remaining members by largest perpendicular distance.
Knee points are intuitively the most attractive points on a Pareto
front when no preference information is available. KnEA pushes the
search toward them at the cost of less uniform front coverage.
Yang, Li, Liu & Zheng 2013 GrEA: many-objective MOEA whose secondary
ranking is a grid-based diversity score instead of crowding distance
or reference vectors.
Each generation:
- NSGA-II-like loop with offspring + non_dominated_sort
- For the splitting front:
- Translate by ideal/nadir; partition objective space into a
(`grid_divisions` per axis) grid
- For every member compute three grid scores:
- GR (grid rank) = sum of grid coordinates (closer to ideal = lower)
- GCD (grid crowding distance) = #neighbors within 1 grid unit (in any axis)
- GCPD (grid coordinate point distance) = max coord - min coord
- Sort F_l ascending by GR, then by GCD, then by GCPD
- Take the top `n - already_selected` survivors
GrEA's grid-based niching is a different lens from NSGA-III's reference
points and RVEA's reference vectors — particularly effective on
non-convex fronts where reference-vector approaches struggle.
Panichella 2019 AGE-MOEA: a many-objective MOEA that *infers* the
front's geometry (its L_p shape, where p = 1 is linear, p = 2 is
spherical, p < 1 is convex etc.) from the current non-dominated set
and uses that estimate to drive both proximity and diversity in
survival selection.
Each generation:
- NSGA-II-like loop: random parent selection + variation + evaluation
- Combine + non_dominated_sort
- Fill front-by-front; for the splitting front:
- Translate by ideal point z*
- Find extreme points by ASF (same as NSGA-III) and intercepts
- Estimate the geometry parameter p by minimizing
\|f − ideal\|_p constancy on the extreme points
- Score every member by survival_score = (proximity_to_ideal) +
(1 / nearest-neighbor distance in the same L_p frame)
- Keep the top scorers
The geometry estimation is the novel contribution; with 3+ objectives
it produces fronts whose spread better matches the true shape than
NSGA-III's reference points (which assume a known geometry).
Rao 2011 TLBO: parameter-free single-objective optimizer for Vec<f64>.
The selling point — uniquely among the metaheuristics we ship — is that
it has NO algorithm-specific hyperparameters: no F, CR, w, c1, c2, σ,
mutation rate, etc. Just population_size and generations.
Each generation has two phases:
- **Teacher phase**: identify the best individual (the 'teacher'). For
every learner, compute a 'mean' learner and try replacing it with a
candidate moved toward the teacher by a random fraction, scaled by
the gap between teacher and (TF · mean), where TF ∈ {1, 2}.
- **Learner phase**: each learner picks a random partner and tries
moving toward the better one of the pair. Only successful moves are
kept.
Single-objective only, Vec<f64> only, bounds enforced via clamping.
Tests cover Sphere1D convergence, deterministic reruns, and panic on
multi-objective.
Lévy-flight perturbation: each variable receives a step drawn from a
heavy-tailed Lévy(α) distribution rather than a Normal. The result is
"mostly small steps with rare big jumps," which gives a more
exploratory mutation than Gaussian without abandoning local search.
Decision type: Vec<f64>, with optional bounds (clamped per-axis if
`bounds` is non-empty). The step is sampled via Mantegna's algorithm
which generates Lévy(α) by combining two Normal samples and taking
the right power, controlled by the tail exponent `alpha` (typical
1.5; 1 is heavy, 2 collapses to Normal).
This is the only genuinely-different mutation kernel from Cuckoo
Search and other Lévy-flight metaheuristics; ship it as a Variation
operator usable from any algorithm rather than as a separate
algorithm.
Adds runners for the five new MO algorithms in both the ZDT1 (2-obj)
and DTLZ2 (3-obj) sections of `examples/compare.rs`. The harness now
side-by-sides 11 multi-/many-objective optimizers (RandomSearch + 10
real ones) on each problem.
Replaces strict Pareto dominance with ε-dominance: A ε-dominates B when
`floor(A_i / ε) ≤ floor(B_i / ε)` for every objective and strictly
less in at least one (minimization frame). The result is a regular
discretization of objective space — at most one archive member per
ε-box — so the front spreads out automatically and the archive size
self-limits without truncation tricks.
Steady-state design: each generation samples one parent from the main
population and one from the ε-archive, applies variation, evaluates
the child, and offers it to both archives. Every member's
ε-coordinates and the box-tie rules are precomputed each insertion.
Tests: produces a front on Schaffer N.1 with reasonable spread,
deterministic reruns, panic on `epsilon[i] <= 0.0` and on
`epsilon.len() != objectives.len()`.
Corne, Jerram, Knowles & Oates 2001: divides objective space into a
hyperbox grid and uses per-box population counts to drive selection
toward sparsely-populated regions.
Each generation:
- Maintain an external archive of non-dominated members
- Build a hyperbox grid (`grid_divisions` per axis on the archive's
current axis ranges); count members per box
- Selection picks two parents by region-based tournament: choose two
random non-empty boxes and take a uniform-random member from the
one with fewer occupants
- Variation produces an offspring; insert into archive, dropping
dominated members and (if archive overflows) the most-crowded
occupant of the most-occupied box
Tests cover non-empty front on Schaffer N.1, deterministic reruns,
and panic on `archive_size == 0`.
Cheng, Jin, Olhofer & Sendhoff 2016 RVEA: many-objective MOEA built
around a fixed set of Das–Dennis reference vectors. Each generation:
- Generate offspring via random parent selection + variation +
evaluation
- Combine population + offspring; translate by ideal point z*
- Associate every member with the reference vector whose angle to
the translated objective vector is smallest
- For each occupied vector, keep the member with the smallest
Angle-Penalized Distance (APD) score; the rest are dropped
- APD = (1 + α(t)·θ_max·γ) · |f − z*| where γ is the angle to the
associated reference and α(t) = (t / t_max)^2 anneals the angle
penalty over the run
This produces well-spread fronts at high objective counts where
Pareto-rank methods (NSGA-II, SPEA2) lose discrimination.
Bader & Zitzler 2011: HypE estimates hypervolume contributions via
Monte Carlo sampling instead of computing them exactly. The point of
the trick is that exact hypervolume becomes prohibitively expensive
beyond ~5 objectives, while MC sampling stays cheap and accurate
enough at any dimension.
Each generation:
- Generate offspring via parent selection + variation + evaluation
- Combine, run non_dominated_sort, fill front-by-front
- For the splitting front, estimate each member's HV contribution
by drawing `n_samples` uniform points in the box [ideal, reference]
and counting how many points are dominated by *exactly* one front
member — that count, divided by n_samples and multiplied by the
box volume, is the member's expected unique HV contribution.
- Drop members one at a time from the splitting front by smallest
estimated contribution.
Public API matches the rest of the MO algorithms (Config + Optimizer).
The reference point is supplied in the config so the user controls
the integration domain. Tests cover non-empty front, deterministic
reruns, and panic on dim-mismatched reference.
Beume, Naujoks & Emmerich 2007: a steady-state MOEA that uses
hypervolume contribution as the secondary survival selection criterion.
Each generation:
- Generate ONE child via parent selection + variation + evaluation.
- Combine population + child, run non_dominated_sort.
- The discarded individual is the worst-front member with the
smallest hypervolume contribution (computed via the new
hypervolume_nd_from_evaluations helper).
Selection-quality is excellent at moderate objective counts (2–4) at
the cost of higher per-step compute (each survival selection requires
N+1 hypervolume evaluations of size ≤ N each). Best paired with a
tightly-bounded objective space — the user supplies a fixed reference
point in the config.
Tests: produces a non-empty front on Schaffer N.1, deterministic
reruns, panic on `population_size == 0`, panic on
`reference_point.len() != objectives.len()`.
Generalizes the existing 2-D hypervolume to arbitrary M ≥ 1 dimensions
using the standard recursive Hypervolume-by-Slicing-Objectives (HSO)
algorithm from While et al. 2006:
- For M = 1: return reference[0] - min(points[0])
- For M = 2: sort by axis 0, sweep accumulating rectangles (matches
hypervolume_2d's existing exact behavior)
- For M ≥ 3: sort by the last axis, peel off slices of increasing
thickness and recursively compute the (M−1)-dimensional HV of each
slice's projected non-dominated subset
Direction-aware: minimization-oriented input is the entry point, so
maximize objectives are negated by the caller via
`ObjectiveSpace::as_minimization` before the recursion runs.
Tested against:
- the existing 2-D analytical case (3 points → area 6)
- a known 3-D unit-cube case (1 point at origin, ref [1,1,1] → 1)
- empty front → 0
- agreement with hypervolume_2d on random 2-D fronts
A substantial README section walking newcomers through choosing an
optimizer. Defines the terminology as it comes up — single- vs multi-
vs many-objective, Pareto front, dominance, multimodality, evaluation
cost — so a reader who has never touched heuristic optimization can
still pick a sensible starting algorithm.
Five-step decision flow:
1. What is the decision?
2. How many objectives?
3. What's the landscape like? (multimodal, smooth, discrete)
4. How expensive is each evaluation?
5. Are there constraints?
Each branch ends with 1–3 algorithm recommendations and a one-line
rationale, plus a compact "quick reference" table at the bottom for
returning users.
Adds runners for HillClimber, SimulatedAnnealing, GeneticAlgorithm,
ParticleSwarm, CmaEs, and Umda to `examples/compare.rs`. Rastrigin
section now compares 8 single-objective optimizers against each other
on a fixed evaluation budget.
The MO sections (ZDT1, DTLZ2) are unchanged for now — MOPSO and IBEA
get added in a follow-up commit so each algorithm's debut shows up
clearly in the harness.
Mühlenbein 1997 UMDA: simplest Estimation-of-Distribution Algorithm for
`Vec<bool>` problems. Each generation:
- Evaluate the current population
- Select the top μ members by fitness
- Estimate per-bit marginal probability p_i = (count of 1s at bit i in
the μ-best) / μ
- Sample population_size new individuals from the resulting product-of-
Bernoullis distribution
Single-objective only. Bit-wise probabilities are clamped to
`[1 / (2 · μ), 1 - 1 / (2 · μ)]` to keep the population from collapsing
to a deterministic single string before convergence is meaningful
(standard Laplace-style smoothing for UMDA).
Tests: solves OneMax (maximize Σ bits) on a 20-bit instance,
deterministic reruns, panic on multi-objective.
Dorigo-style Ant System for permutation problems on a complete graph:
each generation, every ant constructs a tour by probabilistically
picking the next node from those it has not yet visited, weighted by
`τ_ij^α · η_ij^β` where τ is the pheromone level on edge (i, j) and
η is the heuristic desirability (1 / distance, here). After all ants
finish, pheromone evaporates by a factor `(1 - ρ)` and is reinforced
on each ant's tour proportional to that tour's quality.
Decision type is `Vec<usize>` (a permutation of 0..n_cities). The user
supplies a distance matrix and the n_cities is inferred. Single-objective
only (the cost is total tour length, which the Problem evaluates).
Tests build a 5-city ring and verify ACO finds a near-optimal tour,
plus deterministic reruns and panic on multi-objective.
Zitzler & Künzli 2004 IBEA: replaces Pareto-rank + crowding fitness
with a single scalar fitness derived from a binary quality indicator
(here, the additive ε-indicator). Loses no information at three or
more objectives the way crowding distance does.
Algorithm:
- For every (i, j) pair compute I(i, j) = max_k (f_k(i) - f_k(j)) on
minimization-oriented objectives.
- Fitness F(i) = -Σ_{j≠i} exp(-I(j, i) / κ).
- Each generation: combine parents + offspring, iteratively remove the
lowest-F member (cleanly recomputing the contribution of the dropped
member from each surviving member's fitness) until population_size
remain.
- Parent selection: binary tournament on F (higher wins).
Bounds-aware operators recommended (SBX + PolyMut).
Tests: produces a non-empty front on Schaffer N.1, deterministic
reruns, panic on `population_size == 0`.
Coello, Pulido & Lechuga 2004 MOPSO: PSO adapted for multi-objective
optimization via an external Pareto archive used as the source of
swarm leaders.
Each generation:
- Evaluate every particle's current position
- Insert non-dominated members into the archive (using ParetoArchive)
- For each particle, pick a leader from the archive (uniform random
among archive members)
- Update velocity using inertia + cognitive (toward pbest) + social
(toward leader)
- Update positions, clamp to bounds
- Refresh personal bests using Pareto comparison: pbest is replaced
only when the new position dominates it; on non-dominated, keep
with 50/50 random tiebreak
Vec<f64> decisions only. Truncates the archive to `archive_size` via
the existing simple-tail truncation. Tests: produces a non-empty
front on Schaffer N.1, deterministic reruns, panic on
single-objective.
Hansen & Ostermeier 2001 CMA-ES, the canonical real-valued
single-objective stochastic optimizer. Implements the full (μ/μ_w, λ)
update with rank-μ + rank-1 covariance updates and cumulative step-size
adaptation:
- Sample λ offspring from N(mean, σ² · C)
- Select the μ best, weight them, recompute mean
- Update evolution paths p_σ (step size) and p_c (covariance)
- Rank-1 update of C from p_c, plus rank-μ update from selected offspring
- Adapt σ via |p_σ| / E‖N(0,I)‖
Eigendecomposition (used to convert C into its B·D form for sampling
N(0, σ²·C)) goes through the new internal Jacobi helper, recomputed
every `eigen_decomposition_period` generations to amortize cost.
Vec<f64> decisions only. Bounds taken from a `RealBounds` field; mean
and offspring are clamped per dimension. Single-objective only.
Hyperparameters use the standard CMA-ES defaults (μ=λ/2, weights from
Hansen's tutorial, c_σ, c_c, c_1, c_μ, d_σ all formulae from §7.1).
Tests cover: convergence on Sphere1D and 5-D Rosenbrock, deterministic
reruns, panic on multi-objective, panic on `population_size < 4`.
Hand-rolled symmetric-matrix eigendecomposition via the cyclic Jacobi
rotation method. Returns sorted (eigenvalue, eigenvector) pairs in
descending order. Pure f64 row-major `Vec<Vec<f64>>` interface so we
don't pull in nalgebra for one algorithm.
Lives in `src/internal/eigen.rs` (new module). Used by the upcoming
CMA-ES implementation to maintain the covariance matrix's
eigendecomposition each generation. Tested against the standard
2x2 case, the diagonal case, and a known 3x3 result.
Glover 1986 tabu search for single-objective problems. Generic over
decision type — the user supplies a neighbor-generator closure that
produces a finite list of candidate moves from the current incumbent
(e.g. all 2-swaps for a permutation, or N Gaussian-perturbed copies of
a real vector). Each iteration picks the best non-tabu neighbor (with
an aspiration override that lets a tabu move through if it beats the
best-seen-ever incumbent) and adds the chosen move's decision to a
fixed-size FIFO tabu list.
Single-objective only. Tracks the best-seen-ever incumbent across the
run, returned as the result. Generic over the decision `D: Hash + Eq`
so the tabu list can match by full decision (simple and correct;
move-based tabu is left for users to implement themselves via a
custom decision wrapper).
Eberhart & Kennedy 1995 PSO with the standard inertia-weight update:
v[i,t+1] = w·v[i,t] + c1·r1·(pbest[i] - x[i,t]) + c2·r2·(gbest - x[i,t])
x[i,t+1] = clamp(x[i,t] + v[i,t+1], bounds)
Single-objective only, `Vec<f64>` decisions only (PSO's velocity vector
needs a Euclidean structure that doesn't generalize cleanly to bool/perm).
Velocities are clamped to ±(hi - lo) per dim to keep particles from
exploding off into space.
Config exposes the four standard knobs — swarm size, generations,
inertia w, cognitive c1, social c2 — plus a seed. Tests cover
convergence on Sphere1D, deterministic reruns, and panic on
multi-objective.
Canonical generational GA with elitism: each generation runs binary
tournament selection (using `tournament_select_single_objective`) on
the current population, applies the variation operator pair-wise to
produce offspring, evaluates them, then replaces the population while
preserving the top `elitism` members from the previous generation
(elitism prevents fitness regression on a single seed).
Single-objective only. Generic over decision type — pair with
`SimulatedBinaryCrossover + PolynomialMutation` for real-valued,
single-point crossover + bit-flip for binary, etc.
Tests: convergence on Sphere1D, deterministic reruns, panic on
multi-objective, panic on `population_size < 2`, panic on
`elitism > population_size`.
Classic Kirkpatrick et al. 1983 SA: hill climber that also accepts
worse moves with probability `exp(-Δ/T)` where T anneals geometrically
from `initial_temperature` to `final_temperature` over the iteration
count.
Single-objective only. Generic over decision type — works on real
vectors, bool vectors, permutations, anything. Tracks the best-seen
incumbent across the run (not just the last accepted move) so the
result reflects the actual best ever visited, not where the random
walk happened to end.
Tests cover: convergence on Sphere1D under reasonable hyperparameters,
deterministic reruns, panic on multi-objective, panic on
non-positive temperatures.
The simplest possible local search: start from one initializer-sampled
decision, repeatedly mutate it via the variation operator, and keep the
child only when it is strictly better than the current incumbent (with
the standard feasible-beats-infeasible / lower-violation tiebreaks
when relevant).
Single-objective only — panics with a clear message if the problem
exposes more than one objective. Deterministic under a seed. Returns
a population/front of size one (the current incumbent) so it slots
into the comparison harness like any other optimizer.
The Python tune_runtime.py treats the morning boot press and the 13:00
post-lunch re-tap as 'free' and only counts extra warning-phase taps.
That undercounts what the user actually presses each day and breaks
any comparison against a stated 'presses/day' comfort cap.
Updated `simulate_one` to count every press the user makes:
- boot press at workday start (always +1)
- 13:00 re-login press when the workday continues past lunch (+1)
- per-minute Bernoulli warning-phase presses (already counted)
- death-restart press: each time the device transitions running→dead
during workday and the user is at-desk (not at lunch), the user
presses to restart the cycle (warning press and death-restart for
the same cycle are mutually exclusive — extending via warning press
prevents that cycle's death)
With baseline now ~2 presses/day already mandatory, the hinge/cap
shift up too: PRESS_HINGE_LOW = 2.5/d (full reward up to baseline +
half a warning press) and PRESS_COMFORT_CAP = 3.5/d (rejected above).
Output 'Why' bullet now reports the total directly and notes the
component breakdown so the number is interpretable against the
new thresholds.
Tweak the a-posteriori scoring to match the user's stated preferences:
- Reweight: lunch_sleep 30%, after_hours 25%, work_fail 20%,
presses 15% (with hinge below), balance 10%.
- Press term is now a hinge instead of a normalized minimize:
* <= 2 presses/day → score 1.0 (no penalty)
* 2 → 3 presses/day → linear ramp from 1.0 to 0.0
* > 3 presses/day → -inf (excluded; comfort cap)
- New balance term: bonus for longer warning phases. Computed as
min(YA - RA, RA - FRA), saturated at 10 minutes. So a 5/5/X split
scores 0.5, an 8/8/X split scores 0.8, and 10/10/X or wider saturates
at 1.0.
Constants moved to module scope so the printout in main and the
scoring function stay in sync.
Adds an a-posteriori decision step to the jiggly example. After NSGA-III
produces the Pareto front, we apply a weighted-sum score over each
objective normalized to [0, 1] across the front (best→1, worst→0,
direction-aware), and report the top three plus a clear recommendation.
The weights are stated explicitly with rationale, not buried in code:
work_fail 45% — screen sleeping mid-meeting is the worst failure
lunch_sleep 30% — the actual design goal
presses 15% — UX friction the user feels
after_hours 10% — minor, mostly screen burn
This is the standard structure for picking a single answer out of a
Pareto set without losing the front itself: someone with different
weights can read the front and pick differently, but we surface a
specific recommendation with reasoning rather than leaving the user
to stare at 84 incomparable rows. Identical normalization could be
swapped for TOPSIS or knee-point detection later if useful.
Port of `scripts/tune_runtime.py` from ~/Code/jiggly: optimize the four
lifecycle constants of a USB mouse-jiggler firmware so the screen sleeps
during the user's lunch hour rather than failing during work.
The Python script grid-searches against a single composite score that
linearly combines several genuinely conflicting goals — a workaround
for the fact that grid search needs one number to rank by. heuropt has
the actual right tool, so this example is structured as a 4-objective
NSGA-III run that surfaces the Pareto front of legitimate tradeoffs:
1. minimize work-time failures (mean_work_sleep)
2. maximize lunch sleep (mean_lunch)
3. minimize human button presses (mean_presses)
4. minimize after-hours waste (mean_after)
Decision: 4-element `Vec<f64>` for (RT, YA, RA, FRA), continuous-relaxed
and rounded to integer minutes inside `evaluate`. The firmware ordering
constraint YA > RA > FRA > 0 is encoded as `constraint_violation` so
heuropt's feasible-beats-infeasible logic handles it for free.
Solver: NSGA-III with M=4, H=6 → 84 reference points, matching the
population size. Each evaluate runs a 1,000-workday Monte Carlo, so
the example is also a deliberately meaty evaluator that benefits from
`--features parallel`.
Output is in jiggly's native units — RT as Xh00m, thresholds in plain
minutes, sleep durations as Xh00m / Mm, probabilities as percentages —
and contrasts the Pareto front against:
- the four extreme single-axis winners (most lunch / fewest work fails /
fewest presses / least after-hours)
- the firmware's currently-shipping defaults (which sit inside the
front as a balanced compromise)
Two new runners — `zdt1_moead` and `dtlz2_moead` — using the same
SBX + PolyMut variation as the other Pareto-based methods. Reference
divisions chosen so the implied population size is comparable to the
other algorithms in each section (99 → 100 weights for ZDT1; 12 → 91
weights for DTLZ2).
Implementation of Zhang & Li 2007 MOEA/D — the canonical
decomposition-based MOEA. Different paradigm from Pareto-dominance
algorithms: each subproblem is a scalarized single-objective problem
defined by a Das–Dennis weight vector, and subproblems with similar
weight vectors form neighborhoods that share genetic material.
Each generation iterates over every weight vector `i`:
1. Pick two parents uniformly from the T-nearest neighbors of weight i
(T = neighborhood_size).
2. Apply variation, evaluate the child.
3. Update the ideal point z* with the child's objectives.
4. Walk the entire neighborhood: for each j, if the child's
Tchebycheff value g(child | w_j, z*) <= g(current[j] | w_j, z*),
replace current[j] with the child.
Tchebycheff scalarization:
g(f | w, z*) = max_k w_k · |f_k - z*_k|
(With the standard `w_k = 1e-6` floor when a weight is zero, so the
max well-defined.)
Public API:
MoeadConfig {
generations,
reference_divisions, // Das-Dennis H, also fixes population size
neighborhood_size, // T
seed,
}
Moead { config, initializer, variation }
impl<P, I, V> Optimizer<P> for Moead<I, V>
Population size equals the number of weight vectors generated by
das_dennis(num_objectives, reference_divisions). Re-exported from the
prelude. Tests cover non-empty Pareto front, deterministic reruns,
and panic on `reference_divisions` that would yield zero weights.
- nsga3: drop redundant `.into_iter()` in extend call; use
`#[allow(clippy::needless_range_loop)]` on the back-substitution
loop where `j` indexes into the matrix; remove an unneeded
`return` keyword in a closure.
- spea2: switch `pool.extend(x.drain(..))` to `pool.append(&mut x)`.
- examples/compare.rs DTLZ2 evaluator: same `needless_range_loop`
silencer on the inner cosine product loop.
NSGA-III's value over NSGA-II shows up at 3+ objectives, where
crowding distance loses its diversity signal. Adds a third comparison
section to `examples/compare.rs`:
DTLZ2 (3-objective, 12-D, the textbook benchmark for many-objective
algorithms): unit-sphere-octant Pareto front. Compares RandomSearch,
NSGA-II, SPEA2, and NSGA-III on:
- mean distance from front points to the unit sphere
(closed-form: |1 - sqrt(f1² + f2² + f3²)|),
- spacing,
- front size,
- wall-clock ms.
NSGA-III config: H=12 reference divisions (91 reference points,
matching the canonical setup from Deb & Jain 2014).
Also wires NSGA-III into the existing ZDT1 (2-objective) section even
though it's not its sweet spot — useful as a regression check that the
algorithm at least keeps up with NSGA-II on bi-objective problems.
Implementation of Deb & Jain 2014 NSGA-III — the canonical
many-objective MOEA. Replaces NSGA-II's crowding-distance niching
with a structured reference-point niching procedure that scales to
3+ objectives where crowding distance loses its diversity signal.
Each generation:
1. Random parent selection + variation + offspring evaluation, same as
NSGA-II.
2. Combine + non_dominated_sort, fill the next population front-by-
front until the next front would overflow (the splitting front F_l).
3. Survival on F_l uses reference-point niching:
- Translate by the ideal point z* (per-axis min in oriented space).
- Compute extreme points by ASF and intercepts; normalize by
intercepts (with a robust fallback to per-axis range if extreme
points are degenerate).
- Associate every member of the working pool with the closest
reference direction by perpendicular distance.
- Iteratively pick from F_l: prefer the niche with the smallest
count among references that have F_l candidates; if the niche is
empty in the already-selected set, take the closest associated
member by perpendicular distance, otherwise pick uniformly from
the niche.
Public API:
Nsga3Config { population_size, generations, reference_divisions, seed }
Nsga3 { config, initializer, variation }
impl<P, I, V> Optimizer<P> for Nsga3<I, V>
Re-exported from the prelude. Tests cover non-empty Pareto front,
exact final population size, deterministic reruns, and panic on
`population_size == 0`. Uses the existing tests_support problems.
The standard structured weight/reference vector generator for
many-objective MOEAs (NSGA-III, MOEA/D). Generates (H+M-1 choose M-1)
points uniformly distributed on the unit simplex by enumerating all
integer compositions of `divisions` into `num_objectives` parts and
dividing each by `divisions`.
Lives in src/pareto/reference_points.rs. Re-exported from the prelude
as `das_dennis`.
Tests cover: M=2/H=4 → 5 points along the diagonal; M=3/H=12 → 91
points (the canonical NSGA-III 3-objective ref set); each generated
point has exactly M components summing to 1 within float tolerance.
Implementation of Zitzler, Laumanns, Thiele 2001 SPEA2 — the classic
Pareto MOEA built around an explicit external archive of fixed size.
Each generation:
1. Combine the current population and the archive into one pool.
2. For every member, compute strength S(i) = number of others that
member dominates, then raw fitness R(i) = sum of S(j) over members
j that dominate i.
3. Add a density estimator D(i) = 1/(σ_k + 2) where σ_k is the distance
to the k-th nearest neighbor (k = floor(sqrt(|pool|))) in
minimization-oriented objective space.
4. Final fitness F(i) = R(i) + D(i); lower is better.
5. Build the next archive by taking every non-dominated member
(R(i) == 0). If too many, prune by repeatedly removing the member
with the smallest k-th-nearest-neighbor distance. If too few, fill
from the rest sorted by F ascending.
6. Generate the next population by binary tournament on F (lower wins),
then variation, then evaluation.
Public API mirrors the other algorithms:
Spea2Config { population_size, archive_size, generations, seed }
Spea2 { config, initializer, variation }
impl<P, I, V> Optimizer<P> for Spea2<I, V>
Re-exported from the prelude. Tests cover archive size invariants,
non-empty Pareto front on Schaffer N.1, deterministic reruns under
the same seed, and panic on population_size == 0.
A comparison example that runs every applicable optimizer on ZDT1 and
Rastrigin across N seeds and reports mean ± stddev for each quality
metric. Designed so a new algorithm slots in by adding a single runner
function — no harness changes needed.
ZDT1 (multi-objective, dim=30):
Reports hypervolume_2d (against ref point [1.1, 1.1]), spacing, mean
L2 distance to the analytical Pareto front, front size, and wall-clock
ms. RandomSearch, PAES, and NSGA-II all use bounds-aware operators
(RealBounds, BoundedGaussianMutation, SBX+PolyMut) so the Problem
itself stays unclamped — apples-to-apples.
Rastrigin (single-objective, dim=5):
Reports mean ± stddev best objective and ms. RandomSearch, PAES,
NSGA-II (degenerate single-obj case), and DE.
Default budget: 10 seeds × 25,000 evaluations on ZDT1, × 50,000 on
Rastrigin. Run with:
cargo run --release --example compare
- PolynomialMutation::vary: `#[allow(clippy::needless_range_loop)]`
on the per-dimension loop — body indexes both `self.bounds[j]` and
`child[j]` so a range index is the cleanest option.
- Operator tests: replace `x >= lo && x <= hi` with
`(lo..=hi).contains(&x)` per clippy's manual_range_contains lint.
swaits