Heuristic — constructive

▶ Open interactive explainer →

Kohonen Self-Organizing Map
Alias som, kohonen
Type Heuristic — constructive
Complexity O(epochs · N · n) per run

Description

Uses a 1-D ring of neurons — a neural network — that learns a topology-preserving mapping from 2-D city space onto a cyclic ordering. Each neuron holds a position in the same plane as the cities. The ring is initialised as a small circle near the centroid and trained so that it wraps around all the cities; the ordering of neurons on the ring then gives the tour.

Training

City coordinates are first normalised to the unit box [0, 1]² for numerical stability. N = n_cities × neuron_multiplier neurons are placed uniformly on a small circle (radius 0.1) centred at the normalised centroid.

Each epoch:

  1. Anneal learning rate and neighbourhood radius:

    η  = learning_rate × exp(−t / epochs)
σ  = max(radius_fraction × N × exp(−t / epochs), 1.0)

  2. Pick a random city c (with replacement).

  3. Find the Best Matching Unit (BMU) — the neuron nearest to c; ties broken by lowest index.

  4. Pull every neuron toward c weighted by a Gaussian over ring distance:

    h(i) = exp(−ring_dist(i, bmu)² / 2σ²)
neurons[i] += η · h(i) · (c − neurons[i])

    Neurons with h < 0.001 are skipped for performance.

Tour extraction

After training, each city is assigned to its closest neuron (BMU). Cities are sorted by their neuron’s ring index. Collisions (two cities mapping to the same neuron) are resolved by distance to the neuron, with city-array index as the final tie-breaker, guaranteeing a valid Hamiltonian tour regardless of convergence quality.

Options

Field CLI flag Default Range Description
epochs --epochs 100 000 ≥ 1 Number of training iterations
learning_rate --learning_rate 0.8 (0, 1] Initial learning rate η₀
radius_fraction --radius_fraction 0.1 (0, 1] Initial neighbourhood radius as fraction of neuron count
neuron_multiplier --neuron_multiplier 8 ≥ 1 Neuron count = n_cities × multiplier; higher reduces dead neurons

All four fields are also reachable via the REST API’s configs.som or a [som]/[stage.som] TOML table (field names match the table above).

Usage

# standalone
teeline solve som -i ./data/tsplib/berlin52.tsp

# recommended: pipe into local search
teeline pipeline --steps=som,2opt -i ./data/tsplib/berlin52.tsp
teeline pipeline --steps=som,sa   -i ./data/tsplib/berlin52.tsp

# custom training schedule
teeline solve som --epochs=200000 --learning_rate=0.9 -i ./data/tsplib/berlin52.tsp

Per-stage TOML config (via pipeline --config):

[[stage]]
solver = "som"

[stage.som]
epochs        = 50000
learning_rate = 0.9

Notes

  • Coordinate normalisation: city coordinates are normalised to [0, 1]² internally. The tour is decoded by neuron ring index, not neuron position, so no denormalisation is needed.
  • init_tour is ignored: as a constructive solver, SOM builds a tour from scratch and does not benefit from a seed tour; the --init-tour pipeline option has no effect.
  • Dead neurons: neurons that never win the BMU for any city leave gaps in coverage. The default neuron_multiplier = 8 (more neurons than cities) mitigates this. Increase it on highly clustered instances.
  • Premature freezing: if radius_fraction is too small or epochs too few, the ring freezes before reaching all cities. Increase epochs or radius_fraction if the gap is unusually large.
  • Quality: on berlin52, typical gap is ~5–15% standalone. Piping into 2-opt routinely reduces this to ~3–7%.

References

  • Kohonen, T. (1982) — “Self-organized formation of topologically correct feature maps”, Biological Cybernetics 43(1), 59–69.
  • Angéniol, B., de la Croix Vaubois, G., & Le Texier, J.-Y. (1988) — “Self-organizing feature maps and the travelling salesman problem”, Neural Networks 1(4), 289–293.