Island Models

A deme is a sub-population that evolves on its own toolbox, then exchanges individuals with other demes. step_islands runs one evaluate → vary → select step on each deme. A migration callable then moves emigrants.

The contract is easy to get wrong. Each deme toolbox must register vary, select, and evaluate (or evaluate_batch). mate and mutate alone are not enough — wrap them with var_and or var_or and register that as vary.

A complete three-deme OneMax run is the islands example. Creating the list of demes is also in Creating Individuals.

One step

from deap_er import tools

def vary(population):
    return tools.var_and(toolbox, population, cx_prob=0.5, mut_prob=0.2)

toolbox.register("vary", vary)
toolbox.register("select", tools.sel_tournament, contestants=3)
toolbox.register("evaluate", evaluate)

demes = [(toolbox, pop_a), (toolbox, pop_b), (toolbox, pop_c)]
tools.step_islands(demes)

Each pair is (toolbox, population). Populations are modified in place. After the step, select(offspring, len(population)) replaces the deme. Unlike toolboxes are allowed — one deme can use a different vary or evaluate.

Migration

Pass migrate to move individuals after every deme has stepped. mig_ring sends each deme to the next. mig_fully_connected sends along every directed edge. mig_random picks one destination per source. All three take a selection callable and a mig_count:

def migrate(populations):
    tools.mig_ring(populations, mig_count=5, selection=tools.sel_best)

tools.step_islands(demes, migrate=migrate)

Emigrants are placed by object identity. When a replacement operator is omitted, a source individual that would otherwise be aliased into two demes is cloned.

Different exams per deme

When demes evaluate on different case subsets or matrices, a migrant's fitness is stale. Pass eval_keys= so step_islands clears fitness on arrival. island_eval_keys hashes each deme's CaseExam and an optional matrix identity:

keys = tools.island_eval_keys(exams, n_cases=20)
tools.step_islands(demes, migrate=migrate, eval_keys=keys)

Omit eval_keys when every deme scores the same data — migrants keep their fitness.

Persist the list of demes on a Checkpoint. There is no island daemon and no extra process pool; evaluate_batch still owns parallelism. See Multiprocessing and the Algorithms reference.