Performant evolutionary algorithms for Python
Visit the examples to get started, or browse the releases.
Contributions are welcome!
Ruck aims to fill the following criteria:
- Provide high quality, readable implementations of algorithms.
- Be easily extensible and debuggable.
- Performant while maintaining its simplicity.
Ruck has various features that will be expanded upon in time
- π¦ Β Modular evolutionary systems inspired by pytorch lightning
- Helps organise code & arguably looks clean
- π― Β Multi-Objective optimisation support
- Optionally optimised version of NSGA-II if
numba
is installed, over 65x faster than the DEAP equivalent
- Optionally optimised version of NSGA-II if
- π Β Optional multithreading support with
ray
, including helper functions - π Β Factory methods for simple evolutionary algorithms
- π§ͺ Β Various helper functions for selection, mutation and mating
Please use the following citation if you use Ruck in your research:
@Misc{Michlo2021Ruck,
author = {Nathan Juraj Michlo},
title = {Ruck - Performant evolutionary algorithms for Python},
howpublished = {Github},
year = {2021},
url = {https://github.com/nmichlo/ruck}
}
Ruck takes inspiration from PyTorch Lightning's module system. The population creation,
offspring, evaluation and selection steps are all contained within a single module inheriting
from EaModule
. While the training logic and components are separated into its own class.
Members
of a Population
(A list of Members) are intended to be read-only. Modifications should not
be made to members, instead new members should be created with the modified values. This enables us to
easily implement efficient multi-threading, see below!
The trainer automatically constructs HallOfFame
and LogBook
objects which keep track of your
population and offspring. EaModule
provides defaults for get_stats_groups
and get_progress_stats
that can be overridden if you wish to customize the tracked statistics and statistics displayed by tqdm.
import random
import numpy as np
from ruck import *
class OneMaxMinimalModule(EaModule):
"""
Minimal onemax example
- The goal is to flip all the bits of a boolean array to True
- Offspring are generated as bit flipped versions of the previous population
- Selection tournament is performed between the previous population and the offspring
"""
# evaluate unevaluated members according to their values
def evaluate_values(self, values):
return [v.sum() for v in values]
# generate 300 random members of size 100 with 50% bits flipped
def gen_starting_values(self):
return [np.random.random(100) < 0.5 for _ in range(300)]
# randomly flip 5% of the bits of each each member in the population
# the previous population members should never be modified
def generate_offspring(self, population):
return [Member(m.value ^ (np.random.random(m.value.shape) < 0.05)) for m in population]
# selection tournament between population and offspring
def select_population(self, population, offspring):
combined = population + offspring
return [max(random.sample(combined, k=3), key=lambda m: m.fitness) for _ in range(len(population))]
if __name__ == '__main__':
# create and train the population
module = OneMaxMinimalModule()
pop, logbook, halloffame = Trainer(generations=100, progress=True).fit(module)
print('initial stats:', logbook[0])
print('final stats:', logbook[-1])
print('best member:', halloffame.members[0])
Ruck provides various helper functions and implementations of evolutionary algorithms for convenience. The following example makes use of these additional features so that components and behaviour can easily be swapped out.
The three basic evolutionary algorithms provided are based around deap's
default algorithms from deap.algorithms
. These basic evolutionary algorithms can be created from
ruck.functional.make_ea
. We provide the alias ruck.R
for ruck.functional
. R.make_ea
supports
the following modes: simple
, mu_plus_lambda
and mu_comma_lambda
.
Code Example
"""
OneMax serial example based on:
https://github.com/DEAP/deap/blob/master/examples/ga/onemax_numpy.py
"""
import functools
import numpy as np
from ruck import *
class OneMaxModule(EaModule):
def __init__(
self,
population_size: int = 300,
offspring_num: int = None, # offspring_num (lambda) is automatically set to population_size (mu) when `None`
member_size: int = 100,
p_mate: float = 0.5,
p_mutate: float = 0.5,
ea_mode: str = 'simple'
):
# save the arguments to the .hparams property. values are taken from the
# local scope so modifications can be captured if the call to this is delayed.
self.save_hyperparameters()
# implement the required functions for `EaModule`
self.generate_offspring, self.select_population = R.make_ea(
mode=self.hparams.ea_mode,
offspring_num=self.hparams.offspring_num,
mate_fn=R.mate_crossover_1d,
mutate_fn=functools.partial(R.mutate_flip_bit_groups, p=0.05),
select_fn=functools.partial(R.select_tournament, k=3),
p_mate=self.hparams.p_mate,
p_mutate=self.hparams.p_mutate,
)
def evaluate_values(self, values):
return map(np.sum, values)
def gen_starting_values(self) -> Population:
return [
np.random.random(self.hparams.member_size) < 0.5
for i in range(self.hparams.population_size)
]
if __name__ == '__main__':
# create and train the population
module = OneMaxModule(population_size=300, member_size=100)
pop, logbook, halloffame = Trainer(generations=40, progress=True).fit(module)
print('initial stats:', logbook[0])
print('final stats:', logbook[-1])
print('best member:', halloffame.members[0])
If we need to scale up the computational requirements, for example requiring increased member and population sizes, the above serial implementations will soon run into performance problems.
The basic Ruck implementations of various evolutionary algorithms are designed around a map
function that can be swapped out to add multi-threading support. We can easily do this using
ray and we even provide various helper functions that
enhance ray support.
-
We begin by placing member's values into shared memory using ray's read-only object store and the
ray.put
function. These ObjectRef's values point to the originalnp.ndarray
values. When retrieved withray.get
they obtain the original arrays using an efficient zero-copy procedure. This is advantageous over something like Python's multiprocessing module which uses expensive pickle operations to pass data around. -
The second step is to swap out the aforementioned
map
function in the previous example to a multiprocessing equivalent. We useray.remote
along withray.get
, and provide theray_map
function that has the same API as python map, but acceptsray.remote(my_fn).remote
values instead. -
Finally we need to update our
mate
andmutate
functions to handleObjectRef
s, we provide a convenient wrapper to automatically callray.put
on function results so that you can re-use your existing code.
Code Example
"""
OneMax parallel example using ray's object store.
8 bytes * 1_000_000 * 128 members ~= 128 MB of memory to store this population.
This is quite a bit of processing that needs to happen! But using ray
and its object store we can do this efficiently!
"""
from functools import partial
import numpy as np
from ruck import *
from ruck.external.ray import *
class OneMaxRayModule(EaModule):
def __init__(
self,
population_size: int = 300,
offspring_num: int = None, # offspring_num (lambda) is automatically set to population_size (mu) when `None`
member_size: int = 100,
p_mate: float = 0.5,
p_mutate: float = 0.5,
ea_mode: str = 'mu_plus_lambda'
):
self.save_hyperparameters()
# implement the required functions for `EaModule`
self.generate_offspring, self.select_population = R.make_ea(
mode=self.hparams.ea_mode,
offspring_num=self.hparams.offspring_num,
# decorate the functions with `ray_remote_put` which automatically
# `ray.get` arguments that are `ObjectRef`s and `ray.put`s returned results
mate_fn=ray_remote_puts(R.mate_crossover_1d).remote,
mutate_fn=ray_remote_put(R.mutate_flip_bit_groups).remote,
# efficient to compute locally
select_fn=partial(R.select_tournament, k=3),
p_mate=self.hparams.p_mate,
p_mutate=self.hparams.p_mutate,
# ENABLE multiprocessing
map_fn=ray_map,
)
# eval function, we need to cache it on the class to prevent
# multiple calls to ray.remote. We use ray.remote instead of
# ray_remote_put like above because we want the returned values
# not object refs to those values.
self._ray_eval = ray.remote(np.mean).remote
def evaluate_values(self, values):
# values is a list of `ray.ObjectRef`s not `np.ndarray`s
# ray_map automatically converts np.sum to a `ray.remote` function which
# automatically handles `ray.get`ing of `ray.ObjectRef`s passed as arguments
return ray_map(self._ray_eval, values)
def gen_starting_values(self):
# generate objects and place in ray's object store
return [
ray.put(np.random.random(self.hparams.member_size) < 0.5)
for i in range(self.hparams.population_size)
]
if __name__ == '__main__':
# initialize ray to use the specified system resources
ray.init()
# create and train the population
module = OneMaxRayModule(population_size=128, member_size=1_000_000)
pop, logbook, halloffame = Trainer(generations=200, progress=True).fit(module)
print('initial stats:', logbook[0])
print('final stats:', logbook[-1])
print('best member:', halloffame.members[0])