Setup
  mesh: Fig. 3a's hexagonal mesh: 36 tunable units, 72 phase shifters (two a unit), 24 ports; each unit's loss drawn at 0.15 ± 0.05 dB, its passive phases drawn at random
  target: the all-cross state, Fig. 3b's 12 channels, each at the best of 8 frequencies over the response's period
  grid: none: exact solves of the mesh
  solver: 100 meshes a method: the particle swarm of Fig. 3 (108 particles, inertia 0.9 to 0.35 over 1000 operations, cognitive 0.7, social 1.8, 36 moves) and of Table 18 (inertia 0.5, cognitive 0.8, social 1, 26 moves); the genetic algorithm (population 144, 18 parents, 36 mutated within 1.570796 rad, decay 0.5, 22 generations); gradient descent with momentum 0.5 by central differences of 0.0003 rad on 20 meshes, learning rates 50 and 1, to 0.2 in at most 60 iterations; Fig. 3's swarm with 0 to 5 % thermal crosstalk on 20 meshes
  tolerance: Fig. 3's swarm's share above −3 dB within two standard deviations of the difference of two counts of 100; Fig. 3b's channels exactly; the solve against photonoxide's circuit solve to 1e-12

The 36-unit hexagonal mesh of Pérez-López et al. 2020 set to all-cross by a GA and a PSO, without calibration: exact solves, 8 frequencies a period

The mesh: 36 units, 72 phase shifters, 24 ports
  all-cross routes (units crossed): 12–23 (8), 14–21 (6), 10–1 (6), 8–3 (4), 13–6 (6), 15–4 (8), 17–2 (6), 19–24 (4), 11–16 (4), 9–18 (6), 7–20 (8), 5–22 (6)
  Fig. 3b's channels routed                    12   paper 12                     ok
  the solve against photonoxide's circuit solve of the mesh as a netlist, 24 × 24: |ΔS| up to 4e-16
  largest |ΔS|                           0.000000   round-off 0.000000   ± 1e-12 ok
  every unit exactly in cross: OF1 -0.90 dB on average over the 100 meshes, their loss alone

Particle swarm, as in Fig. 3: 108 particles, cognitive 0.7, social 1.8, the inertia from 0.9 to 0.35 over the first 1000 operations
  79 of 100 meshes above −3 dB (paper 76.66 %); OF1 -2.51 dB on average; 3996 operations
  PSO's share above −3 dB                0.790000   paper, 2σ of 2 counts 0.766600   ± 1.196e-1 ok

Particle swarm, Suppl. Table 18: 108 particles, inertia 0.5, cognitive 0.8, social 1
  57 of 100 meshes above −3 dB (paper 70 %); OF1 -3.15 dB on average; 2916 operations
  2.0 standard deviations of the difference from the paper's 70 %: printed, not checked

Genetic algorithm: 144 genomes, the best 18 kept as parents, one-point crossover, 36 phases moved within ±0.5π·e^(−0.5 t). Printed, not checked: the paper doesn't say what t counts
  t the generation:
  4 of 100 meshes above −3 dB (paper 95 %); OF1 -4.27 dB on average; 2916 operations
  t the operations in thousands:
  99 of 100 meshes above −3 dB (paper 95 %); OF1 -2.06 dB on average; 2916 operations

Gradient descent with momentum 0.5, central differences of 0.0003 rad (145 operations an iteration), from uniform phases until CF < 0.2 or 60 iterations, 20 meshes. Printed, not checked: the paper's 5 to 6 iterations at a learning rate of 50 hang on how its CF is scaled, which it doesn't say
  learning rate 50: CF < 0.2 on 0 of 20 meshes; OF1 -11.75 dB on average where they stopped
  learning rate 1: CF < 0.2 on 19 of 20 meshes in a median of 18 iterations (2610 operations); OF1 -1.68 dB on average where they stopped

The particle swarm of Fig. 3 with crosstalk from 0 to 5 % between every pair of shifters (Fig. 4's), 20 meshes. Printed, not checked: the paper has no all-cross run with it
  12 of 20 meshes above −3 dB (paper none); OF1 -2.99 dB on average; 3996 operations
all 3 within tolerance of the paper
