Skip to content

Getting Started

Installation

This library can be installed with:

pip install deap-er

or if you're using the uv package manager:

uv add deap-er

The optional Numba compile backend for columnar GP is an extra:

pip install deap-er[numba]

uv add deap-er --extra numba

Namespaces

The functionality of this library is divided into the following namespaces:

  • deap_er — Toolbox, Fitness, Checkpoint, and the creator module.
  • tools — algorithms (ea_*, MAP-Elites, islands, optional n_evals=), operators (lexicase, SMS-EMOA, MOEA/D, AGE-MOEA-II, mixed-gene variation, DE, constraint-dominance), CMA strategies (boxed, separable, restarting, MO-CMA), records (logbook, hall of fame, MAP-Elites archives), utilities (EvalCache, spawn_rng, affine_scale, case_errors, semantic helpers), and benchmarks.
  • gp — prefix-tree GP (loosely typed, strongly typed, ADFs), SlimGP, growing language, memetic / affine writeback, and columnar kits with opcode / Numba backends.

These namespaces can be imported with:

from deap_er import Checkpoint, Fitness, Toolbox, creator, gp, tools

First program

Register operators, build a population, run ea_simple. This OneMax run maximizes the number of ones in a 20-bit string:

from deap_er import Fitness, Toolbox, creator, tools

tools.rng.seed(1234)

creator.create_type("FitnessMax", Fitness, weights=(1.0,))
creator.create_type("Individual", list, fitness=creator.FitnessMax)

toolbox = Toolbox()
toolbox.register("attr_bool", tools.rng.randint, 0, 1)
toolbox.register(
    "individual", tools.init_repeat, creator.Individual, toolbox.attr_bool, 20
)
toolbox.register("population", tools.init_repeat, list, toolbox.individual)
toolbox.register("mate", tools.cx_two_point)
toolbox.register("mutate", tools.mut_flip_bit, mut_prob=0.05)
toolbox.register("select", tools.sel_tournament, contestants=3)
toolbox.register("evaluate", lambda ind: (sum(ind),))

pop = toolbox.population(size=60)
hof = tools.HallOfFame(1)
tools.ea_simple(
    toolbox, pop, generations=20, cx_prob=0.5, mut_prob=0.2, hof=hof, verbose=True
)
print(hof[0], hof[0].fitness.values)

A longer walkthrough of the same problem is the One Max example.

Published benchmarks

The tools barrel ships common test functions as bm_* callables. Register one as evaluate instead of writing a fitness function:

toolbox.register("evaluate", tools.bm_sphere)      # unimodal
toolbox.register("evaluate", tools.bm_rastrigin)   # multimodal
toolbox.register("evaluate", tools.bm_zdt_1)       # two-objective

bm_sphere and bm_rastrigin return a one-float tuple. bm_zdt_1 returns two objectives — pair it with a multi-objective selector.

Where next

Tutorials, in a useful reading order: