https://pasaopasen.github.io/geneticalgorithm2/

geneticalgorithm2 (from DPEA) is the supported advanced optimized fork of non-supported package geneticalgorithm of Ryan (Mohammad) Solgi

About
Installation
Updates information
Working process
Examples for beginner
U should know these features
Examples pretty collection
Popular questions

About

geneticalgorithm2 is very flexible and highly optimized Python library for implementing classic genetic-algorithm (GA).

Features of this package:

written on pure python
extremely fast
no hard dependencies (only numpy primary)
easy to run: no need to perform long task-setup process
easy to logging, reach support of flexible callbacks
many built-in plotting functions
many built-in cases of crossover, mutation and selection
support of integer, boolean and real (continuous/discrete) variables types
support of mixed types of variables
support of classic, elitist and studEA genetic algorithm combinations
support of revolutions and duplicates utilization
reach support of customization

Installation

Install this package with standard light dependencies to use the base functional.

pip install geneticalgorithm2

Install this package with full dependencies to use all provided functional including plotting and built-in parallelism tools.

pip install geneticalgorithm2[full]

Updates information

Future

duplicates removing and revolutions will be moved to MiddleCallbacks and removed as alone run() parameters
new stop criteria callbacks (min std, max functions evaluations)
vartype will support strings like iiiiibbf

TODO firstly

Remove old style mentions from README

6.9.2 bugfixes

6.9.1 refactor

Finally move function_timeout and function to run() method and deprecate its usage in init()
function is not mandatory to be non-empty
reduce documentation duplicates

6.9.0 reborn

recreate the repository without excess heavy files materials
host the code documentation
rename geneticalgorithm2 class to GeneticAlgorithm2
substantial package architecture refactor
add more docstrings

6.8.7 minor update

some code refactor
fixes:
- ensure the directory of generation file exists on save

6.8.6 minor update

small package installation update: add pip install geneticalgorithm2[full] version
small refactor

6.8.5 minor update

update OppOpPopInit 2.0.0->2.0.1
set default function_timeout to None which means no use of function time checking
remove joblib and func_timeout from necessary dependencies

6.8.4 minor update

a bit of refactor
little optimizations
add empty field fill_children(pop_matrix, parents_count) to geneticalgorithm2 class to specify children creating behavior (what is the most intensive part of algorithm after optimizing func calculations), see this

6.8.3 types update

much more type hints

6.8.2 patch

for printing info
fix logic: now population is always sorted before going to callbacks

6.8.1 patch

printing progress bar to 'stderr' or 'stdout' or None (disable) by choice (progress_bar_stream argument of run()), deprecated disable_progress_bar
little speed up
new geneticalgorithm2.vectorized_set_function set function, which can be faster for big populations

6.8.0 minor update

remove crossover_probability model parameter because of it has no sense to exist (and 1.0 value is better than others, take a look at results). This parameter came from geneticalgorithm old package and did`t change before.

6.7.7 refactor

change some behavior about parents selection

6.7.6 bug fix

fix some bug of variable_type=='bool'
some refactor of progress bar
add some dependencies to setup.py

6.7.5 refactor

shorter progress bar (length can be controlled by setting PROGRESS_BAR_LEN field of geneticalgorithm2 class)
shorter logic of run(), more informative output

6.7.4 bug fix

bug fix

6.7.3 speed up

refactor to make run() method faster

6.7.2 little update

better flexible logic for report, take a look
removed show mean parameter from model.plot_result and now model reports only best score by default, not average and so on (u can specify if u wanna report average too)
plot_several_lines useful function

6.7.1 patch

changes according to new OppOpPopInit version

6.7.0 minor update (new features)

add mutation_discrete_type and mutation_discrete_probability parameters in model. It controls mutation behavior for discrete (integer) variables and works like mutation_type and mutation_probability work for continuous (real) variables. Take a look at algorithm parameters

6.6.2 patch (speed up)

fix and speed up mutation

6.6.1 patch

removed unnecessary dependencies

6.6.0 minor update (refactoring)

deprecated variable_type_mixed, now use variable_type for mixed optimization too
deprecated output_dict, now it's better object with name result
refactor of big part of tests
refactor of README

6.5.1 patch

replace collections.Sequence with collections.abc.Sequence, now it should work for python3.10+

6.5.0 minor update (refactoring)

another form of data object using with middle callbacks (MiddleCallbackData dataclass instead of dictionary)
type hints for callbacks module

6.4.1 patch (bug fix)

fix bug setting attribute to algorithm parameters (in middle callbacks)

6.4.0 minor update (refactoring)

new valid forms for start_generation; now it's valid to use
- None
- str path to saved generation
- dictionary with structure {'variables': variables/None, 'scores': scores/None}
- Generation object: Generation(variables = variables, scores = scores)
- np.ndarray with shape (samples, dim) for only population or (samples, dim+1) for concatenated population and score (scores is the last matrix column)
- tuple(np.ndarray/None, np.ndarray/None) for variables and scores
here variables is 2D numpy array with shape (samples, dim), scores is 1D numpy array with scores (function values) for each sample; here and here u can see examples of using these valid forms

6.3.0 minor update (refactoring)

type hints for entire part of functions
new valid forms for function parameters (now u don't need to use numpy arrays everywhere)
AlgorithmParams class for base GA algorithm parameters (instead of dictionary)
Generation class for saving/loading/returning generation (instead of dictionary)

All that classes are collected in file. To maintain backward compatibility, AlgorithmParams and Generation classes have dictionary-like interface for getting fields: u can use object.field or object['field'] notations.

Working process

Main algorithm structure

Pre-process: making inner functions depends on params, making/loading start population

while True:

    if reason to stop (time is elapsed / no progress / generation count is reached / min value is reached):
        break


    select parents to crossover from last population and put them to new population:
        select (elit count) best samples
        select (parents count - elit count) random samples (by selection function)

    create (total samples count - parents count) children (samples from selected parents) and put them to new population:
        while not all children are created:
            select 2 random parents
            make child1, child2 from them using crossover
            mutate child1 by mutation (model.mut)
            mutate child2 by middle mutation (model.mut_middle)
            put children to new population
    
    remove duplicates, make revolutions, sort population by scores
    use callbacks, use middle callbacks

Post-process: plotting results, saving

Optimization process components

Function to minimize

The goal of the optimization process is to find the minimum of the given function (1D array) -> float where the function argument is a vector of some values in different dimensions.

If u want to find the maximum, use this idea:

opt_func = lambda arr: -func(arr)

#
# ... find global min of opt_func
#

opt_minimum=opt_func(best value)
maximum = -opt_minimum

Also it is possible and highly recommended to create and use a vectorized version of this function called set_function (2D array) -> (1D array) which transforms several samples matrix to samples scores vector by one call. Using this way u can speed up calculations or set up more complex tasks optimization

Optimization space

The function rates 1D arrays (vectors) where each component (dimension) means something u program it to mean. Each dimension has its bound ([min; max] cut) and variable type (real/discrete).

Advice. Genetic algorithms work much faster and efficient for discrete tasks. If high precision is not required u can split any real dimension to many discrete values (for instance, [1.1, 1.2, 1.25, 1.44]) and try to optimize indexes of the given array which are converted to real values inside function itself.

Algorithm parameters

There are a number of hyperparameters u can probe to optimize including population size and selection/crossover/mutation types.

Samples constructors

There are several ways to create new testing samples from zero when u start with empty population or when u need new samples after duplicates removing and revolutions.

Callbacks

Now the package supports 2 different types of highly customized callbacks:

simple callbacks
middle callbacks

How to run

Firstly, u should import needed packages.

All available (but not always necessary) imports are:

import numpy as np

# the only one required import
from geneticalgorithm2 import GeneticAlgorithm2 as ga  # for creating and running optimization model

from geneticalgorithm2 import Generation, AlgorithmParams  # classes for comfortable parameters setting and getting

from geneticalgorithm2 import Crossover, Mutations, Selection  # classes for specific mutation and crossover behavior

from geneticalgorithm2 import get_population_initializer  # for creating better start population

from geneticalgorithm2 import np_lru_cache  # for cache function (if u want)

from geneticalgorithm2 import plot_pop_scores  # for plotting population scores, if u want

from geneticalgorithm2 import Callbacks  # simple callbacks (will be deprecated)

from geneticalgorithm2 import Actions, ActionConditions, MiddleCallbacks  # middle callbacks

Next step: define the function to minimize:

def function(X: np.ndarray) -> float: # X as 1d-numpy array
    return np.sum(X**2) + X.mean() + X.min() + X[0]*X[2] # some float result

Also u should create the bounds for each variable (if exist) such as:

var_bound = np.array([[0,10]]*3) # 2D numpy array with shape (dim, 2)

# also u can use Sequence of Tuples (from version 6.3.0)
var_bound = [
    (0, 10),
    (0, 10),
    (0, 10)
]

Important. U don't need to use variable boundaries only if variable type of each variable is boolean. This case will be automatically converted to discrete variables with bounds (0, 1).

After that u create a GeneticAlgorithm2 (was imported early as ga) object:

model = ga( 
    dimension = 3, 
    variable_type='real', 
    variable_boundaries = var_bound,
    algorithm_parameters={
        'max_num_iteration': None,
        'population_size':100,
        'mutation_probability': 0.1,
        'mutation_discrete_probability': None,
        'elit_ratio': 0.01,
        'parents_portion': 0.3,
        'crossover_type':'uniform',
        'mutation_type': 'uniform_by_center',
        'mutation_discrete_type': 'uniform_discrete',
        'selection_type': 'roulette',
        'max_iteration_without_improv':None
    }
)

Note: it is not mandatory to write all possible algorithm_parameters, here it is done only to show u defaults. Also u can use AlgorithmParams (with typehints and docstrings) class instead of dicts:

algorithm_parameters=AlgorithmParams(
    max_num_iteration=None,
    population_size=100,
    mutation_probability=0.1,
    mutation_discrete_probability=None,
    elit_ratio=0.01,
    parents_portion=0.3,
    crossover_type='uniform',
    mutation_type='uniform_by_center',
    mutation_discrete_type='uniform_discrete',
    selection_type='roulette',
    max_iteration_without_improv=None
)

Run the search method:

# all of this parameters are default
result = model.run(
    no_plot = False, 
    progress_bar_stream = 'stdout',
    disable_printing = False,

    function=function,
    function_timeout=None,

    set_function = None, 
    apply_function_to_parents = False, 
    start_generation = None,
    studEA = False,
    mutation_indexes = None,

    init_creator = None,
    init_oppositors = None,
    duplicates_oppositor = None,
    remove_duplicates_generation_step = None,
    revolution_oppositor = None,
    revolution_after_stagnation_step = None,
    revolution_part = 0.3,
    
    population_initializer = Population_initializer(select_best_of = 1, local_optimization_step = 'never', local_optimizer = None),
    
    stop_when_reached = None,
    callbacks = [],
    middle_callbacks = [],
    time_limit_secs = None, 
    save_last_generation_as = None,
    seed = None
)

# best candidate
print(result.variable)

# best score
print(result.score)

# last generation
print(result.last_generation)

Constructor parameters

Have a look at https://pasaopasen.github.io/geneticalgorithm2/geneticalgorithm2/geneticalgorithm2.html#GeneticAlgorithm2.__init__

Genetic algorithm's parameters

AlgorithmParams object

The parameters of GA are defined as a dictionary or AlgorithmParams object: https://pasaopasen.github.io/geneticalgorithm2/geneticalgorithm2/data_types/algorithm_params.html

To get the global default params use code:

params = ga.default_params

To get actual parameters of an existing model use code:

params = model.param

Parameters of algorithm

Methods and Properties of model

Have a look at https://pasaopasen.github.io/geneticalgorithm2/geneticalgorithm2/geneticalgorithm2.html#GeneticAlgorithm2.run

Examples for beginner

A minimal example

Assume we want to find a set of X = (x1, x2, x3) that minimizes function f(X) = x1 + x2 + x3 where X can be any real number in [0, 10].

This is a trivial problem and we already know that the answer is X = (0, 0, 0) where f(X) = 0.

We just use this simple example to show how to implement it with geneticalgorithm2. First we import geneticalgorithm2 and numpy. Next, we define function f which we want to minimize and the boundaries of the decision variables. Then simply geneticalgorithm2 is called to solve the defined optimization problem as follows:

import numpy as np
from geneticalgorithm2 import GeneticAlgorithm2 as ga


def f(X):
    return np.sum(X)


varbound = [[0, 10]] * 3

model = ga(dimension=3, variable_type='real', variable_boundaries=varbound)

model.run(function=f)

If you run the code, you should see a progress bar that shows the progress of the genetic algorithm (GA) and then the solution, objective function value and the convergence curve as follows:

Also we can access to the best answer of the defined optimization problem found by GA as a dictionary and a report of the progress of the genetic algorithm. To do so we complete the code as follows:

convergence = model.report

solution = model.result

The simple example with integer variables

Considering the problem given in the simple example above. Now assume all variables are integers. So x1, x2, x3 can be any integers in [0, 10]. In this case the code is as the following:

import numpy as np
from geneticalgorithm2 import GeneticAlgorithm2 as ga


def f(X):
    return np.sum(X)

varbound = [[0, 10]] * 3

model = ga(dimension=3, variable_type='int', variable_boundaries=varbound)

model.run(function=f)

So, as it is seen the only difference is that for variable_type we use string 'int'.

The simple example with Boolean variables

Considering the problem given in the simple example above. Now assume all variables are boolean instead of real or integer. So X can be either zero or one. Also instead of three let's have 30 variables. In this case the code is as the following:

import numpy as np
from geneticalgorithm2 import GeneticAlgorithm2 as ga


def f(X):
    return np.sum(X)

model = ga(dimension=30, variable_type='bool')

model.run(function=f)

Note for variable_type we use string 'bool' when all variables are boolean.
Note that when variable_type equal 'bool' there is no need for variable_boundaries to be defined.

The simple example with mixed variables

Considering the problem given in the the simple example above where we want to minimize f(X) = x1 + x2 + x3. Now assume x1 is a real (continuous) variable in [0.5; 1.5], x2 is an integer variable in [1;100], and x3 is a boolean variable that can be either zero or one. We already know that the answer is X = (0.5, 1, 0) where f(X) = 1.5. We implement geneticalgorithm2 as the following:

import numpy as np
from geneticalgorithm2 import GeneticAlgorithm2 as ga


def f(X):
    return np.sum(X)


varbound = [[0.5, 1.5], [1, 100], [0, 1]]
vartype = ('real', 'int', 'int')
model = ga(dimension=3, variable_type=vartype, variable_boundaries=varbound)

model.run(function=f)

Optimization problems with constraints

In all above examples, the optimization problem was unconstrained. Now consider that we want to minimize f(X) = x1+x2+x3 where X is a set of real variables in [0; 10]. Also we have an extra constraint so that sum of x1 and x2 is equal or greater than 2. The minimum of f(X) is 2. In such a case, a trick is to define penalty function. Hence we use the code below:

import numpy as np
from geneticalgorithm2 import GeneticAlgorithm2 as ga


def f(X):
    pen = 0
    if X[0] + X[1] < 2:
        pen = 500 + 1000 * (2 - X[0] - X[1])
    return np.sum(X) + pen


varbound = [[0, 10]] * 3

model = ga(dimension=3, variable_type='real', variable_boundaries=varbound)

model.run(function=f)

As seen above we add a penalty to the objective function whenever the constraint is not met.

Some hints about how to define a penalty function:

Usually you may use a constant greater than the maximum possible value of the objective function if the maximum is known or if we have a guess of that. Here the highest possible value of our function is 300 (i.e. if all variables were 10, f(X)=300). So I chose a constant of 500. So, if a trial solution is not in the feasible region even though its objective function may be small, the penalized objective function (fitness function) is worse than any feasible solution.
Use a coefficient big enough and multiply that by the amount of violation. This helps the algorithm learn how to approach feasible domain.
How to define penalty function usually influences the convergence rate of an evolutionary algorithm. In my book on metaheuristics and evolutionary algorithms you can learn more about that.
Finally after you solved the problem test the solution to see if boundaries are met. If the solution does not meet constraints, it shows that a bigger penalty is required. However, in problems where optimum is exactly on the boundary of the feasible region (or very close to the constraints) which is common in some kinds of problems, a very strict and big penalty may prevent the genetic algorithm to approach the optimal region. In such a case designing an appropriate penalty function might be more challenging. Actually what we have to do is to design a penalty function that let the algorithm searches unfeasible domain while finally converge to a feasible solution. Hence you may need more sophisticated penalty functions. But in most cases the above formulation work fairly well.

Middle example: select fixed count of objects from set

For some task u need to think a lot and create good specific crossover or mutation functions. For example, take a look at this problem:

    From set like X = {x1, x2, x3, ..., xn} u should select only k objects which get the best function value

U can do it using this code:

import numpy as np
from geneticalgorithm2 import GeneticAlgorithm2 as ga

subset_size = 20  # how many objects we can choose

objects_count = 100  # how many objects are in set

my_set = np.random.random(objects_count) * 10 - 5  # set values


# minimized function
def f(X):
    return abs(np.mean(my_set[X == 1]) - np.median(my_set[X == 1]))


# initialize start generation and params

N = 1000  # size of population
start_generation = np.zeros((N, objects_count))
indexes = np.arange(0, objects_count, dtype=np.int8)  # indexes of variables

for i in range(N):
    inds = np.random.choice(indexes, subset_size, replace=False)
    start_generation[i, inds] = 1


def my_crossover(parent_a, parent_b):
    a_indexes = set(indexes[parent_a == 1])
    b_indexes = set(indexes[parent_b == 1])

    intersect = a_indexes.intersection(b_indexes)  # elements in both parents
    a_only = a_indexes - intersect  # elements only in 'a' parent
    b_only = b_indexes - intersect

    child_inds = np.array(list(a_only) + list(b_only), dtype=np.int8)
    np.random.shuffle(child_inds)  # mix

    children = np.zeros((2, parent_a.size))
    if intersect:
        children[:, np.array(list(intersect))] = 1
    children[0, child_inds[:int(child_inds.size / 2)]] = 1
    children[1, child_inds[int(child_inds.size / 2):]] = 1

    return children[0, :], children[1, :]


model = ga(
    dimension=objects_count,
    variable_type='bool',
    algorithm_parameters={
        'max_num_iteration': 500,
        'mutation_probability': 0,  # no mutation, just crossover
        'elit_ratio': 0.05,
        'parents_portion': 0.3,
        'crossover_type': my_crossover,
        'max_iteration_without_improv': 20
    }
)

model.run(
    function=f,
    no_plot=False, 
    start_generation=(start_generation, None)
)

U should know these features

Available crossovers

For two example parents (one with ones and one with zeros) next crossovers will give same children (examples):

one_point:

0	0	0	0	0	0	0	0	1	1	1	1	1	1	1
1	1	1	1	1	1	1	1	0	0	0	0	0	0	0

two_point:

1	1	0	0	0	0	0	0	0	0	0	0	1	1	1
0	0	1	1	1	1	1	1	1	1	1	1	0	0	0

uniform:

1	1	1	0	1	1	0	0	0	0	0	1	0	0	0
0	0	0	1	0	0	1	1	1	1	1	0	1	1	1

uniform_window:

1	1	1	1	1	1	1	1	1	0	0	0	1	1	1
0	0	0	0	0	0	0	0	0	1	1	1	0	0	0

shuffle:

0	0	0	1	1	1	1	0	0	1	1	1	0	1	0
1	1	1	0	0	0	0	1	1	0	0	0	1	0	1

segment:

0	1	1	0	0	1	0	1	0	0	1	0	0	1	1
1	0	0	1	1	0	1	0	1	1	0	1	1	0	0

arithmetic:

0.13	0.13	0.13	0.13	0.13	0.13	0.13	0.13	0.13	0.13	0.13	0.13	0.13	0.13	0.13
0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87	0.87

mixed:

0.63	0.84	1.1	0.73	0.67	-0.19	0.3	0.72	-0.18	0.61	0.84	1.14	1.36	-0.37	-0.19
0.51	0.58	0.43	0.42	0.55	0.49	0.57	0.48	0.46	0.56	0.56	0.54	0.44	0.51	0.4

Available selections

Function timeout

geneticalgorithm2 inherited several features from geneticalgorithm package sush as that if the given function does not provide any output before timeout, the algorithm would be terminated and raise the appropriate error.

In such a case make sure the given function works correctly (i.e. there is no infinite loop in the given function). Also if the given function takes more than 10 seconds to complete the work make sure to increase function_timeout in arguments.

Standard GA vs. Elitist GA

The convergence curve of an elitist genetic algorithm is always non-increasing. So, the best ever found solution is equal to the best solution of the last iteration. However, the convergence curve of a standard genetic algorithm is different. If elit_ratio is zero geneticalgorithm2 implements a standard GA. The output of geneticalgorithm2 for standard GA is the best ever found solution not the solution of the last iteration. The difference between the convergence curve of standard GA and elitist GA is shown below:

Standard crossover vs. stud EA crossover

Stud EA is the idea of using crossover always with best object. So one of two parents is always the best object of population. It can help us in a lot of tasks!

Creating better start population

There is get_population_initializer(select_best_of = 4, local_optimization_step = 'never', local_optimizer = None) function for creating start population creators. Take a look at its docs

Select best N of kN

This little option can help u especially with multimodal tasks.

Do local optimization

We can apply some local optimization on start generation before starting GA search. It can be some gradient descent or hill climbing and so on. Also we can apply it before selection best objects (on entire population) or after (on best part of population) and so forth.

In next example I'm using my DiscreteHillClimbing algorithm for local optimization my discrete task:

import numpy as np
import matplotlib.pyplot as plt

from DiscreteHillClimbing import Hill_Climbing_descent

from geneticalgorithm2 import GeneticAlgorithm2 as ga
from geneticalgorithm2 import get_population_initializer


def f(arr):
    arr2 = arr / 25
    return -np.sum(arr2 * np.sin(np.sqrt(np.abs(arr2)))) ** 5 + np.sum(np.abs(arr2)) ** 2


iterations = 100

varbound = [[-100, 100]] * 15

available_values = [np.arange(-100, 101)] * 15

my_local_optimizer = lambda arr, score: Hill_Climbing_descent(
    function=f, available_predictors_values=available_values,
    max_function_evals=50, start_solution=arr
)

model = ga(
    dimension=varbound.shape[0],
    variable_type='int',
    variable_boundaries=varbound,
    algorithm_parameters={
        'max_num_iteration': iterations,
        'population_size': 400
    }
)

for time in ('before_select', 'after_select', 'never'):
    model.run(
        function=f
        no_plot=True,
        population_initializer=get_population_initializer(
            select_best_of=3,
            local_optimization_step=time,
            local_optimizer=my_local_optimizer
        )
    )

    plt.plot(model.report, label=f"local optimization time = '{time}'")

plt.xlabel('Generation')
plt.ylabel('Minimized function (40 simulations average)')
plt.title('Selection best N object before running GA')
plt.legend()

Optimization with oppositions

Also u can create start population with oppositions. See example of code

Revolutions

U can create revolutions in your population after some stagnation steps. It really can help u for some tasks. See example

Duplicates removing

If u remove duplicates each k generations, u can speed up the optimization process (example)

Cache

It can be useful for run-speed to use cache with some discrete tasks. For this u can import np_lru_cache decorator and use it like here:

import np_lru_cache

@np_lru_cache(maxsize=some_size)
def minimized_func(arr):
    # code
    return result

#
# run
#    algorithm
#


# don't forget to clear cache
minimized_func.cache_clear()

Report checker

Basically the model checks best population score (minimal score of generation) each generation and saves it to report field. Actually this sequence of numbers u see in big part of plots. This behavior is needed for several parts and u cannot disable it. But if u want to report some other metric without using callbacks, there is highly simple and fast way.

After creating model but before running run() u need to append ur logic to model.checked_reports field. Take a look at example:

import numpy as np

from geneticalgorithm2 import GeneticAlgorithm2 as ga
from geneticalgorithm2 import plot_several_lines


def f(X):
    return 50 * np.sum(X) - np.sum(np.sqrt(X) * np.sin(X))


dim = 25
varbound = [[0, 10]] * dim

model = ga(
    dimension=dim,
    variable_type='real', 
    variable_boundaries=varbound,
    algorithm_parameters={
        'max_num_iteration': 600
    }
)

# here model exists and has checked_reports field
# now u can append any functions to report

model.checked_reports.extend(
    [
        ('report_average', np.mean),
        ('report_25', lambda arr: np.quantile(arr, 0.25)),
        ('report_50', np.median)
    ]
)

# run optimization process
model.run(
    function=f,
    no_plot=False
)

# now u have not only model.report but model.report_25 and so on

# plot reports
names = [name for name, _ in model.checked_reports[::-1]]
plot_several_lines(
    lines=[getattr(model, name) for name in names],
    colors=('green', 'black', 'red', 'blue'),
    labels=['median value', '25% quantile', 'mean of population', 'best pop score'],
    linewidths=(1, 1.5, 1, 2),
    title="Several custom reports with base reports",
    save_as='./output/report.png'
)

As u see, u should append tuple (name of report, func to evaluate report) to model.checked_report. It's highly recommended to start this name with report_ (e. g. report_my_median). And the function u use will get 1D-numpy sorted array of population scores.

Middle callbacks

There is an amazing way to control optimization process using MiddleCallbacks class. Just learn next logic:

u can use several MiddleCallbacks callbacks as list at middle_callbacks parameter in run() method
each middle callback is the pair of action and condition functions
condition(data) (Callable[[MiddleCallbackData], bool]) function gets data object (dataclass MiddleCallbackData from version 6.5.0) about primary model parameters and makes logical decision about applying action function
action(data) (Callable[[MiddleCallbackData],MiddleCallbackData]) function modifies data objects as u need -- and model will be modified by new data

data object is the structure with several parameters u can modify:

 data = MiddleCallbackData(
     last_generation=Generation.from_pop_matrix(pop),
     current_generation=t,
     report_list=self.report,

     mutation_prob=self.prob_mut,
     crossover_prob=self.prob_cross,
     mutation=self.real_mutation,
     crossover=self.crossover,
     selection=self.selection,

     current_stagnation=counter,
     max_stagnation=self.max_stagnations,

     parents_portion=self.param.parents_portion,
     elit_ratio=self.param.elit_ratio,

     set_function=self.set_function
 )

So, the action function gets data objects and returns data object.

It's very simple to create your own action and condition functions. But there are several popular functions contained in Actions and ActionConditions classes:

actions:
- Stop() -- just stop optimization process
- ReduceMutationProb(reduce_coef = 0.9) -- reduce mutation probability
- ChangeRandomCrossover(available_crossovers: Sequence[Callable[[np.ndarray, np.ndarray], Tuple[np.ndarray, np.ndarray]]]) -- change another (random) crossover from list of crossovers
- ChangeRandomSelection(available_selections: Sequence[Callable[[np.ndarray, int], np.ndarray]])
- ChangeRandomMutation(available_mutations: Sequence[Callable[[float, float, float], float]])
- RemoveDuplicates(oppositor = None, creator = None, converter = None); see doc
- CopyBest(by_indexes) -- copies best population object values (from dimensions in by_indexes) to all population
- PlotPopulationScores(title_pattern = lambda data: f"Generation {data['current_generation']}", save_as_name_pattern = None) -- plot population scores; needs 2 functions like data->string for title and file name (to save)
conditions:
- ActionConditions.EachGen(generation_step = 10) -- do action each generation_step generations
- ActionConditions.Always() do action each generations, equals to ActionConditions.EachGen(1)
- ActionConditions.AfterStagnation(stagnation_generations = 50) -- do action after stagnation_generations stagnation generations
- ActionConditions.Several(list_of_conditions) -- do action if all conditions in list are true

To combine action and condition to callback, just use MiddleCallbacks.UniversalCallback(action, condition) methods.

There are also next high-level useful callbacks:

MiddleCallbacks.ReduceMutationGen(reduce_coef = 0.9, min_mutation = 0.005, reduce_each_generation = 50, reload_each_generation = 500)
MiddleCallbacks.GeneDiversityStats(step_generations_for_plotting:int = 10) -- plots some duplicates statistics each gen (example)

See code example

How to compare efficiency of several versions of GA optimization

To compare efficiency of several versions of GA optimization (such as several values of several hyperparameters or including/excepting some actions like oppositions) u should make some count of simulations and compare results using some statistical test. I have realized this logic here

Hints on how to adjust genetic algorithm's parameters (from `geneticalgorithm` package)

In general the performance of a genetic algorithm or any evolutionary algorithm depends on its parameters. Parameter setting of an evolutionary algorithm is important. Usually these parameters are adjusted based on experience and by conducting a sensitivity analysis. It is impossible to provide a general guideline to parameter setting but the suggestions provided below may help:

Number of iterations: Select a max_num_iterations sufficiently large; otherwise the reported solution may not be satisfactory. On the other hand selecting a very large number of iterations increases the run time significantly. So this is actually a compromise between the accuracy you want and the time and computational cost you spend.
Population size: Given a constant number of functional evaluations (max_num_iterations times population_size) I would select smaller population size and greater iterations. However, a very small choice of population size is also deteriorative. For most problems I would select a population size of 100 unless the dimension of the problem is very large that needs a bigger population size.
elit_ratio: Although having few elites is usually a good idea and may increase the rate of convergence in some problems, having too many elites in the population may cause the algorithm to easily trap in a local optima. I would usually select only one elite in most cases. Elitism is not always necessary and in some problems may even be deteriorative.
mutation_probability: This is a parameter you may need to adjust more than the other ones. Its appropriate value heavily depends on the problem. Sometimes we may select mutation_probability as small as 0.01 (i.e. 1 percent) and sometimes even as large as 0.5 (i.e. 50 percent) or even larger. In general if the genetic algorithm trapped in a local optimum increasing the mutation probability may help. On the other hand if the algorithm suffers from stagnation reducing the mutation probability may be effective. However, this rule of thumb is not always true.
parents_portion: If parents_portion set zero, it means that the whole of the population is filled with the newly generated solutions. On the other hand having this parameter equals 1 (i.e. 100 percent) means no new solution is generated and the algorithm would just repeat the previous values without any change which is not meaningful and effective obviously. Anything between these two may work. The exact value depends on the problem.
crossover_type: Depends on the problem. I would usually use uniform crossover. But testing the other ones in your problem is recommended.
max_iteration_without_improv: This is a parameter that I recommend being used cautiously. If this parameter is too small then the algorithm may stop while it trapped in a local optimum. So make sure you select a sufficiently large criteria to provide enough time for the algorithm to progress and to avoid immature convergence.

Finally to make sure that the parameter setting is fine, we usually should run the algorithm for several times and if convergence curves of all runs converged to the same objective function value we may accept that solution as the optimum. The number of runs depends but usually five or ten runs is prevalent. Notice that in some problems several possible set of variables produces the same objective function value. When we study the convergence of a genetic algorithm we compare the objective function values not the decision variables.

How to get maximum speed

Don't use plotting

result = model.run(
    no_plot = True, 
)

Don't use progress bar

result = model.run(
    progress_bar_stream = None,
)

Try to use faster optimizing function

Try to speed up your optimizing function using Numpy, Numba or Cython. If u can, write your own set_function (function which applies to whole population samples matrix) with cython optimizations, parallelism and so.

Specify custom optimized `mutation`, `crossover`, `selection`

Write faster implementations for model methods mut, mut_middle, crossover, selection and set them before running optimization process:

model.mut = custom_mut
model.crossover = custom_crossover

model.run(...)

Specify `fill_children` method

From version 6.8.4 there is fill_children model method:

self.fill_children: Optional[Callable[[array2D, int], None]] = None

It is empty and does nothing; but if u specify it, u can get huge speed up at very intensive algorithm part. Take a look at main algo structure. There is a part with creating children from parents, this part is the most intensive because it uses python loops, calls sampling, crossover and mutations at each iteration. Using fill_children, u can rewrite this logic in your manner to speed up.

Suppose u have new population matrix pop (type np.float64, shape (population_size, dim_count)) where first parents_count rows are selected parents, next rows are filled by random, so inside fill_children method u should fill last population_size - parents_count rows (children) by using some your logic. Expected (but not mandatory) logic like this:

for k in range(self.parents_count, self.population_size, 2):

    r1, r2 = get_parents_inds()  # get 2 random parents indexes from [0, parents_count)

    pvar1 = pop[r1]
    pvar2 = pop[r2]

    ch1, ch2 = self.crossover(pvar1, pvar2)  # crossover

    # mutations
    ch1 = self.mut(ch1)
    ch2 = self.mut_middle(ch2, pvar1, pvar2)

    # put to population
    pop[k] = ch1
    pop[k+1] = ch2

Example. In one task I use this algorithm many times (100 000 generations total), so the speed matters. Every sample item is the index of element in other array there, so ith sample element is always integer value from cut [0, end[i]]. I use uniform crossover and uniform mutation (work perfect for this task). So I specified creating children logic for this task using cython.

Content of file fill_children.pyx:

#!python
#cython: language_level=3

import numpy as np

cimport numpy as np

np.import_array()

cimport cython

import math
import random

@cython.boundscheck(False)
@cython.wraparound(False)
def fill_children(
    np.ndarray[np.float64_t, ndim=2] pop,  # samples are integers but always float64 type
    int parents_count, # count of already done parents

    float mut_prob,  # mutation probability
    np.ndarray[np.uint8_t, ndim=1] ends  # max elements for each dimension (min elements are 0)
):

    cdef:
        Py_ssize_t i, k, population_size = pop.shape[0], dim_count = pop.shape[1], r1, r2

        float v1, v2, tmp
        np.ndarray[np.float64_t, ndim=1] cross, mut, mut_middle

    # making 2 children at each iteration
    for k in range(parents_count, population_size, 2): # C loop, not Python
        
        #
        # 2 random parents (fast implementation)
        #

        r1 = random.randrange(parents_count)
        r2 = random.randrange(parents_count)
        if r1 == r2:
            while r1 == r2:  # C loop!
               r2 = random.randrange(parents_count) 

        #
        # I always need these 3 random probs sequences, so the fastest way to obtain them is np.random.random
        #
        cross = np.random.random(dim_count)  # crossover probabilities for each dimension
        mut = np.random.random(dim_count)
        mut_middle = np.random.random(dim_count)

        for i in range(dim_count):  # C loop for each dimension
            v1 = pop[r1, i]  # first parent value
            v2 = pop[r2, i]  # second parent value

            if cross[i] < 0.5:  # random swap (uniform crossover), copy otherwise
                tmp = v2
                v2 = v1
                v1 = tmp

            if mut[i] < mut_prob:  # random mutation for first child
                # fastest way to get random integer from [0, ends[i]]
                # random.random() calls not always but only on mut[i] < mut_prob
                v1 = math.floor(random.random() * (ends[i] + 1))

            if mut_middle[i] < mut_prob: # mut_middle for second
                tmp = random.random()
                if v1 < v2:
                    v2 = v1 + math.floor(tmp * (v2 - v1 + 1))  # integer from [v1, v2], v1 < v2
                elif v1 > v2:
                    v2 = v2 + math.floor(tmp * (v1 - v2 + 1)) # integer from [v2, v1], v2 < v1
                else:
                    v2 = math.floor(tmp * (ends[i] + 1))

            #
            # put values to children in array
            #
            pop[k, i] = v1
            pop[k + 1, i] = v2

After compilation this file I can call it from python file to use inside GA:

mut_prob = param['mutation_probability']

def fill_children(pop: array2D, parents_count: int):
    """wrapper on fill_children.fill_children with putting local variables mut_prob, ends"""
    return fill_children.fill_children(
        pop, parents_count, mut_prob, ends
    )

model.fill_children = fill_children

model.run(...)

Examples pretty collection

Optimization test functions

Here there is the implementation of geneticalgorithm2 for some benchmark problems. Test functions are got from my OptimizationTestFunctions package.

The code for optimizations process is same for each function and is contained in file.

Using GA in reinforcement learning

See example of using GA optimization with keras neural networks for solving OpenGym tasks.

Better example is OpenGym using cost2fitness and geneticalgorithm2 where I use also my cost2fitness package for fast forward propagation

Using GA with image reconstruction by polygons

Links:

https://www.kaggle.com/demetrypascal/fork-of-imagereconstruction-with-geneticalgorithm2
https://www.kaggle.com/demetrypascal/imagereconstructionpolygons-with-geneticalgorithm2

Files

README.md

Latest commit

History