swaits/heuropt
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs: correct disconnected-front and sequencing guidance from compare results

Browse Source 
The `compare` harness contradicts two recommendations in the decision
trees:

- "Disconnected or non-convex front -> AGE-MOEA, KnEA, IBEA" had it
  backwards. Added KnEA to the ZDT3 table (the disconnected-front
  benchmark) so the claim is actually exercised: AGE-MOEA and KnEA
  finish *last and second-last*; IBEA wins, MOEA/D and NSGA-II follow.
  The trees now split "disconnected" from "non-convex contiguous",
  lead disconnected with IBEA, and note the geometry-aware methods
  trail when the front is in pieces.
- The book filed Simulated Annealing on permutations as a "one-decision
  baseline" and led the JSS row with GA. On the harness SA *wins* the
  FT06 job-shop table and ties for the TSP optimum; SA/Tabu edge out
  the GA. Reframed SA/Tabu as strong sequencing methods.

Also regenerated examples/compare-results.md for the new ZDT3 row.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
swaits
2026-05-14 08:43:55 -06:00
co-authored by Claude Opus 4.7
parent 7004a572c8
commit a9d72b94f4
4 changed files with 63 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+13 -6
View File
@@ -412,10 +412,13 @@ START
│ │ → GA with BitFlipMutation +
│ │ a bit-string crossover
│ │
│ ├─ Decision is Vec<usize> (permutation, e.g., TSP)
│ │ → Ant Colony (with a distance matrix)
│ │ → Tabu Search (with your own neighbor function)
│ │ → Simulated Annealing with SwapMutation
│ ├─ Decision is Vec<usize> (permutation: TSP, JSS, …)
│ │ → Ant Colony (TSP, with a distance matrix)
│ │ → Simulated Annealing / Tabu Search (strong on
│ │ sequencing — they win the harness TSP and JSS
│ │ tables — you supply the neighbour move)
│ │ → GA + permutation toolkit (ERX for TSP-shaped
│ │ instances)
│ │
│ └─ Custom decision type (a struct, a tree, …)
│ → Simulated Annealing or Hill Climber
@@ -448,10 +451,14 @@ START
│ ├─ Want decomposition / weight-vector style
│ │ → MOEA/D (very fast per generation, scales well)
│ │
│ ├─ Disconnected or non-convex front
│ ├─ Disconnected front (separate arcs, e.g. ZDT3)
│ │ → IBEA (wins ZDT3 hypervolume on the harness;
│ │ MOEA/D and NSGA-II follow. Geometry-aware
│ │ methods trail when the front is in pieces)
│ │
│ ├─ Non-convex but *contiguous* front
│ │ → AGE-MOEA (estimates front geometry adaptively)
│ │ → KnEA (favors knee points)
│ │ → IBEA
│ │
│ ├─ Want region-based diversity
│ │ → PESA-II (grid hyperboxes drive selection)
docs/book/src/choosing-an-algorithm.md
+15 -5
View File
@@ -104,9 +104,9 @@ optimizer or with the problem).
| `Vec<bool>` | [GA][GeneticAlgorithm] + [`BitFlipMutation`] | When bit interactions matter. |
| `Vec<usize>` (permutation) | [Ant Colony][AntColonyTsp] | TSP-style with a distance matrix. |
| `Vec<usize>` (permutation) | [GA][GeneticAlgorithm] + [`ShuffledPermutation`] + [`OrderCrossover`] + [`InversionMutation`] | Generic permutation GA; use [`EdgeRecombinationCrossover`] for TSP-shaped instances. |
| `Vec<usize>` (JSS multiset) | [GA][GeneticAlgorithm] + [`ShuffledMultisetPermutation`] + local POX + [`InversionMutation`] | Operation-string encoding; see [Optimize a permutation](./cookbook/permutation.md). |
| `Vec<usize>` (permutation) | [Simulated Annealing][SimulatedAnnealing] + [`SwapMutation`] | One-decision baseline. |
| `Vec<usize>` or custom | [Tabu Search][TabuSearch] | You supply the neighbor function. |
| `Vec<usize>` (JSS multiset) | [Simulated Annealing][SimulatedAnnealing] / [Tabu Search][TabuSearch] with [`InsertionMutation`], or [GA][GeneticAlgorithm] + [`ShuffledMultisetPermutation`] + local POX | Operation-string encoding. On the FT06 harness the local-search pair edges out the GA — see [Optimize a permutation](./cookbook/permutation.md). |
| `Vec<usize>` (permutation) | [Simulated Annealing][SimulatedAnnealing] + [`InversionMutation`] | Strong on sequencing, not just a baseline — wins the harness's FT06 job-shop table and ties for the TSP optimum. |
| `Vec<usize>` or custom | [Tabu Search][TabuSearch] | You supply the neighbor function; consistently near the top on the TSP and JSS tables. |
| Custom struct | [Simulated Annealing][SimulatedAnnealing] / [Hill Climber][HillClimber] | With your own `Variation` impl. |
heuropt's permutation operator toolkit covers four crossovers
@@ -161,14 +161,24 @@ objectives.
### Disconnected or non-convex front
A *disconnected* front (separate arcs, like ZDT3) and a *non-convex but
contiguous* front are different problems — don't conflate them.
For a **disconnected** front, [IBEA][Ibea] is the clear pick: on the
harness it wins ZDT3 — the disconnected-front benchmark — outright on
hypervolume, with [MOEA/D][Moead] and [NSGA-II][Nsga2] close behind.
Counter-intuitively the geometry-aware methods below *trail* here:
estimating a single front geometry or chasing knee points doesn't help
when the front is in pieces (on ZDT3, AGE-MOEA and KnEA finish last).
For a **non-convex but contiguous** front:
[AGE-MOEA][AgeMoea] estimates the front geometry adaptively (the L_p
parameter `p` is fit from data each generation).
[KnEA][Knea] favors knee points — the regions of the front where small
gains in one objective cost large losses in another.
[IBEA][Ibea] also handles disconnected fronts well.
### Region-based diversity
[PESA-II][PesaII] uses grid hyperboxes to drive selection — divide the
examples/compare-results.md
+11 -4
View File
@@ -60,10 +60,17 @@ dominated gaps between them.
| algorithm | hypervolume ↑ | spacing ↓ | front | ms |
|---|---|---|---|---|
| **IBEA** | **126.2072 ± 1.2280** | 0.0164 ± 0.0036 | 48 | 131 |
| MOEA/D | 125.2413 ± 2.1647 | 0.0198 ± 0.0043 | 92 | 27 |
| NSGA-II | 123.1826 ± 1.5829 | **0.0092 ± 0.0020** | 98 | 40 |
| AGE-MOEA | 119.5132 ± 1.2732 | 0.0136 ± 0.0023 | 90 | 171 |
| **IBEA** | **126.2072 ± 1.2280** | 0.0164 ± 0.0036 | 48 | 126 |
| MOEA/D | 125.2413 ± 2.1647 | 0.0198 ± 0.0043 | 92 | 26 |
| NSGA-II | 123.1826 ± 1.5829 | **0.0092 ± 0.0020** | 98 | 39 |
| AGE-MOEA | 119.5132 ± 1.2732 | 0.0136 ± 0.0023 | 90 | 170 |
| KnEA | 117.2180 ± 0.7027 | 0.0147 ± 0.0049 | 79 | 32 |
The **geometry-aware methods finish last** on the disconnected front:
AGE-MOEA and KnEA both trail the dominance- and decomposition-based
methods. Estimating a single front geometry — or chasing knee points —
doesn't help when the front is in pieces; IBEA's indicator-based
selection wins here.
## DTLZ2 (3-obj, dim=12, 30000 evals/run × 10 seeds)
examples/compare.rs
+24
View File
@@ -1574,6 +1574,29 @@ fn zdt3_age_moea(seed: u64) -> MoRun {
}
}
fn zdt3_knea(seed: u64) -> MoRun {
let problem = zdt3_problem();
let bounds = vec![(0.0, 1.0); ZDT3_DIM];
let initializer = RealBounds::new(bounds.clone());
let variation = CompositeVariation {
crossover: SimulatedBinaryCrossover::new(bounds.clone(), 15.0, 0.5),
mutation: PolynomialMutation::new(bounds, 20.0, 1.0 / ZDT3_DIM as f64),
};
let pop = 100;
let config = KneaConfig {
population_size: pop,
generations: ZDT3_BUDGET / pop,
seed,
};
let mut opt = Knea::new(config, initializer, variation);
let t0 = Instant::now();
let result = opt.run(&problem);
MoRun {
front: result.pareto_front,
wall_ms: t0.elapsed().as_millis(),
}
}
// -----------------------------------------------------------------------------
// DTLZ1 runners (curated many-obj subset)
// -----------------------------------------------------------------------------
@@ -1998,6 +2021,7 @@ fn run_zdt3_comparison() {
("MOEA/D", zdt3_moead),
("IBEA", zdt3_ibea),
("AGE-MOEA", zdt3_age_moea),
("KnEA", zdt3_knea),
];
let mut rows: Vec<(f64, Vec<String>)> = Vec::new();
for (name, runner) in runners {