Artificial Bee Colony

artificial_bee_colony() keeps a colony of solutions and improves them with three kinds of bees. Use it when you want a population method with a single knob for how long to persist on a solution, and no crossover to design.

from pymetaheuristics.artificial_bee_colony import artificial_bee_colony

abc = artificial_bee_colony(
    knapsack,
    stop=max_iterations(50),
    rng=1,
    neighbor=bit_flip_neighbor,
    colony_size=10,
    limit=20,
)
print(abc.best_solution, abc.best_value)

Each iteration:

1. Employed bees: every solution tries one neighbor and keeps it only if it is better.
2. Onlooker bees: colony_size more tries, each on a solution picked by a tournament between two random ones, so good solutions get more tries.
3. Scouts: a solution that failed more than limit times in a row is replaced by a fresh one from problem.generate.

The neighbor is the same neighbor(solution, rng) function that simulated annealing uses, so any neighborhood works. An infeasible neighbor is redrawn up to max_neighbor_tries times (default 100), and counts as a failed try if none is feasible.

One iteration costs about 2 * colony_size evaluations (more when scouts fire), and stop is checked once per iteration, so max_evaluations may be overshot by up to one iteration.

Tuning

colony_size and limit have generic defaults (20 and 50). A smaller colony gives each solution more tries for a fixed budget, and a smaller limit restarts stuck solutions sooner. For continuous problems the step of the neighbor (for example gaussian_neighbor(sigma=...)) matters as much.

For the equations, the differences from the original algorithm and measured results, see Artificial Bee Colony in depth.

Recap

• You supply a neighbor move; colony_size and limit are the knobs.
• Each iteration costs about 2 * colony_size evaluations.

Next: read the results.