Release Date: 15 April 2020

1. The pygad.GA class accepts a new argument named fitness_func which accepts a function to be used for calculating the fitness values for the solutions. This allows the project to be customized to any problem by building the right fitness function.

Release Date: 4 May 2020

1. The pygad.GA attributes are moved from the class scope to the instance scope.
2. Raising an exception for incorrect values of the passed parameters.
3. Two new parameters are added to the pygad.GA class constructor (init_range_low and init_range_high) allowing the user to customize the range from which the genes values in the initial population are selected.
4. The code object __code__ of the passed fitness function is checked to ensure it has the right number of parameters.

Release Date: 13 May 2020

1. The fitness function accepts a new argument named sol_idx representing the index of the solution within the population.
2. A new parameter to the pygad.GA class constructor named initial_population is supported to allow the user to use a custom initial population to be used by the genetic algorithm. If not None, then the passed population will be used. If None, then the genetic algorithm will create the initial population using the sol_per_pop and num_genes parameters.
3. The parameters sol_per_pop and num_genes are optional and set to None by default.
4. A new parameter named callback_generation is introduced in the pygad.GA class constructor. It accepts a function with a single parameter representing the pygad.GA class instance. This function is called after each generation. This helps the user to do post-processing or debugging operations after each generation.

Release Date: 14 May 2020

1. The best_solution() method in the pygad.GA class returns a new output representing the index of the best solution within the population. Now, it returns a total of 3 outputs and their order is: best solution, best solution fitness, and best solution index. Here is an example:

solution, solution_fitness, solution_idx = ga_instance.best_solution()
print("Parameters of the best solution :", solution)
print("Fitness value of the best solution :", solution_fitness, "\n")
print("Index of the best solution :", solution_idx, "\n")

1. A new attribute named best_solution_generation is added to the instances of the pygad.GA class. it holds the generation number at which the best solution is reached. It is only assigned the generation number after the run() method completes. Otherwise, its value is -1.

   Example:

print("Best solution reached after {best_solution_generation} generations.".format(best_solution_generation=ga_instance.best_solution_generation))

1. The best_solution_fitness attribute is renamed to best_solutions_fitness (plural solution).
2. Mutation is applied independently for the genes.

Release Date: 17 May 2020

1. Adding 2 extra modules (pygad.nn and pygad.gann) for building and training neural networks with the genetic algorithm.

Release Date: 18 May 2020

1. The initial value of the generations_completed attribute of instances from the pygad.GA class is 0 rather than None.
2. An optional bool parameter named mutation_by_replacement is added to the constructor of the pygad.GA class. It works only when the selected type of mutation is random (mutation_type="random"). In this case, setting mutation_by_replacement=True means replace the gene by the randomly generated value. If False, then it has no effect and random mutation works by adding the random value to the gene. This parameter should be used when the gene falls within a fixed range and its value must not go out of this range. Here are some examples:

Assume there is a gene with the value 0.5.

If mutation_type="random" and mutation_by_replacement=False, then the generated random value (e.g. 0.1) will be added to the gene value. The new gene value is 0.5+0.1=0.6.

If mutation_type="random" and mutation_by_replacement=True, then the generated random value (e.g. 0.1) will replace the gene value. The new gene value is 0.1.

1. None value could be assigned to the mutation_type and crossover_type parameters of the pygad.GA class constructor. When None, this means the step is bypassed and has no action.

Release date: 1 June 2020

1. A new module named pygad.cnn is supported for building convolutional neural networks.
2. A new module named pygad.gacnn is supported for training convolutional neural networks using the genetic algorithm.
3. The pygad.plot_result() method has 3 optional parameters named title, xlabel, and ylabel to customize the plot title, x-axis label, and y-axis label, respectively.
4. The pygad.nn module supports the softmax activation function.
5. The name of the pygad.nn.predict_outputs() function is changed to pygad.nn.predict().
6. The name of the pygad.nn.train_network() function is changed to pygad.nn.train().

Release date: 5 July 2020

1. A new parameter named delay_after_gen is added which accepts a non-negative number specifying the time in seconds to wait after a generation completes and before going to the next generation. It defaults to 0.0 which means no delay after the generation.
2. The passed function to the callback_generation parameter of the pygad.GA class constructor can terminate the execution of the genetic algorithm if it returns the string stop. This causes the run() method to stop.

One important use case for that feature is to stop the genetic algorithm when a condition is met before passing though all the generations. The user may assigned a value of 100 to the num_generations parameter of the pygad.GA class constructor. Assuming that at generation 50, for example, a condition is met and the user wants to stop the execution before waiting the remaining 50 generations. To do that, just make the function passed to the callback_generation parameter to return the string stop.

Here is an example of a function to be passed to the callback_generation parameter which stops the execution if the fitness value 70 is reached. The value 70 might be the best possible fitness value. After being reached, then there is no need to pass through more generations because no further improvement is possible.

def func_generation(ga_instance):
 if ga_instance.best_solution()[1] >= 70:
     return "stop"

Release date: 19 July 2020

1. 2 new optional parameters added to the constructor of the pygad.GA class which are crossover_probability and mutation_probability.

   While applying the crossover operation, each parent has a random value generated between 0.0 and 1.0. If this random value is less than or equal to the value assigned to the crossover_probability parameter, then the parent is selected for the crossover operation.

   For the mutation operation, a random value between 0.0 and 1.0 is generated for each gene in the solution. If this value is less than or equal to the value assigned to the mutation_probability, then this gene is selected for mutation.

2. A new optional parameter named linewidth is added to the plot_result() method to specify the width of the curve in the plot. It defaults to 3.0.

3. Previously, the indices of the genes selected for mutation was randomly generated once for all solutions within the generation. Currently, the genes' indices are randomly generated for each solution in the population. If the population has 4 solutions, the indices are randomly generated 4 times inside the single generation, 1 time for each solution.

4. Previously, the position of the point(s) for the single-point and two-points crossover was(were) randomly selected once for all solutions within the generation. Currently, the position(s) is(are) randomly selected for each solution in the population. If the population has 4 solutions, the position(s) is(are) randomly generated 4 times inside the single generation, 1 time for each solution.

5. A new optional parameter named gene_space as added to the pygad.GA class constructor. It is used to specify the possible values for each gene in case the user wants to restrict the gene values. It is useful if the gene space is restricted to a certain range or to discrete values. For more information, check the More about the ``gene_space` Parameter <https://pygad.readthedocs.io/en/latest/README_pygad_ReadTheDocs.html#more-about-the-gene-space-parameter>`__ section. Thanks to Prof. Tamer A. Farrag for requesting this useful feature.

Release Date: 6 August 2020

1. A bug fix in assigning the value to the initial_population parameter.
2. A new parameter named gene_type is added to control the gene type. It can be either int or float. It has an effect only when the parameter gene_space is None.
3. 7 new parameters that accept callback functions: on_start, on_fitness, on_parents, on_crossover, on_mutation, on_generation, and on_stop.

Release Date: 11 September 2020

1. The learning_rate parameter in the pygad.nn.train() function defaults to 0.01.
2. Added support of building neural networks for regression using the new parameter named problem_type. It is added as a parameter to both pygad.nn.train() and pygad.nn.predict() functions. The value of this parameter can be either classification or regression to define the problem type. It defaults to classification.
3. The activation function for a layer can be set to the string "None" to refer that there is no activation function at this layer. As a result, the supported values for the activation function are "sigmoid", "relu", "softmax", and "None".

To build a regression network using the pygad.nn module, just do the following:

1. Set the problem_type parameter in the pygad.nn.train() and pygad.nn.predict() functions to the string "regression".
2. Set the activation function for the output layer to the string "None". This sets no limits on the range of the outputs as it will be from -infinity to +infinity. If you are sure that all outputs will be nonnegative values, then use the ReLU function.

Check the documentation of the pygad.nn module for an example that builds a neural network for regression. The regression example is also available at this GitHub project: https://github.com/ahmedfgad/NumPyANN

To build and train a regression network using the pygad.gann module, do the following:

1. Set the problem_type parameter in the pygad.nn.train() and pygad.nn.predict() functions to the string "regression".
2. Set the output_activation parameter in the constructor of the pygad.gann.GANN class to "None".

Check the documentation of the pygad.gann module for an example that builds and trains a neural network for regression. The regression example is also available at this GitHub project: https://github.com/ahmedfgad/NeuralGenetic

To build a classification network, either ignore the problem_type parameter or set it to "classification" (default value). In this case, the activation function of the last layer can be set to any type (e.g. softmax).

Release Date: 11 September 2020

1. A bug fix when the problem_type argument is set to regression.

Release Date: 14 September 2020

1. Bug fix to support building and training regression neural networks with multiple outputs.

Release Date: 20 September 2020

1. Support of a new module named kerasga so that the Keras models can be trained by the genetic algorithm using PyGAD.

Release Date: 3 October 2020

1. Bug fix in applying the crossover operation when the crossover_probability parameter is used. Thanks to Eng. Hamada Kassem, Research and Teaching Assistant, Construction Engineering and Management, Faculty of Engineering, Alexandria University, Egypt.

Release Date: 06 December 2020

1. The fitness values of the initial population are considered in the best_solutions_fitness attribute.
2. An optional parameter named save_best_solutions is added. It defaults to False. When it is True, then the best solution after each generation is saved into an attribute named best_solutions. If False, then no solutions are saved and the best_solutions attribute will be empty.
3. Scattered crossover is supported. To use it, assign the crossover_type parameter the value "scattered".
4. NumPy arrays are now supported by the gene_space parameter.
5. The following parameters (gene_type, crossover_probability, mutation_probability, delay_after_gen) can be assigned to a numeric value of any of these data types: int, float, numpy.int, numpy.int8, numpy.int16, numpy.int32, numpy.int64, numpy.float, numpy.float16, numpy.float32, or numpy.float64.

Release Date: 03 January 2021

1. Support of a new module pygad.torchga to train PyTorch models using PyGAD. Check its documentation.
2. Support of adaptive mutation where the mutation rate is determined by the fitness value of each solution. Read the Adaptive Mutation section for more details. Also, read this paper: Libelli, S. Marsili, and P. Alba. "Adaptive mutation in genetic algorithms." Soft computing 4.2 (2000): 76-80.
3. Before the run() method completes or exits, the fitness value of the best solution in the current population is appended to the best_solution_fitness list attribute. Note that the fitness value of the best solution in the initial population is already saved at the beginning of the list. So, the fitness value of the best solution is saved before the genetic algorithm starts and after it ends.
4. When the parameter parent_selection_type is set to sss (steady-state selection), then a warning message is printed if the value of the keep_parents parameter is set to 0.
5. More validations to the user input parameters.
6. The default value of the mutation_percent_genes is set to the string "default" rather than the integer 10. This change helps to know whether the user explicitly passed a value to the mutation_percent_genes parameter or it is left to its default one. The "default" value is later translated into the integer 10.
7. The mutation_percent_genes parameter is no longer accepting the value 0. It must be >0 and <=100.
8. The built-in warnings module is used to show warning messages rather than just using the print() function.
9. A new bool parameter called suppress_warnings is added to the constructor of the pygad.GA class. It allows the user to control whether the warning messages are printed or not. It defaults to False which means the messages are printed.
10. A helper method called adaptive_mutation_population_fitness() is created to calculate the average fitness value used in adaptive mutation to filter the solutions.
11. The best_solution() method accepts a new optional parameter called pop_fitness. It accepts a list of the fitness values of the solutions in the population. If None, then the cal_pop_fitness() method is called to calculate the fitness values of the population.

Release Date: 10 January 2021

1. In the gene_space parameter, any None value (regardless of its index or axis), is replaced by a randomly generated number based on the 3 parameters init_range_low, init_range_high, and gene_type. So, the None value in [..., None, ...] or [..., [..., None, ...], ...] are replaced with random values. This gives more freedom in building the space of values for the genes.
2. All the numbers passed to the gene_space parameter are casted to the type specified in the gene_type parameter.
3. The numpy.uint data type is supported for the parameters that accept integer values.
4. In the pygad.kerasga module, the model_weights_as_vector() function uses the trainable attribute of the model's layers to only return the trainable weights in the network. So, only the trainable layers with their trainable attribute set to True (trainable=True), which is the default value, have their weights evolved. All non-trainable layers with the trainable attribute set to False (trainable=False) will not be evolved. Thanks to Prof. Tamer A. Farrag for pointing about that at GitHub.

Release Date: 15 January 2021

1. A bug fix when save_best_solutions=True. Refer to this issue for more information: ahmedfgad#25

Release Date: 16 February 2021

1. In the gene_space argument, the user can use a dictionary to specify the lower and upper limits of the gene. This dictionary must have only 2 items with keys low and high to specify the low and high limits of the gene, respectively. This way, PyGAD takes care of not exceeding the value limits of the gene. For a problem with only 2 genes, then using gene_space=[{'low': 1, 'high': 5}, {'low': 0.2, 'high': 0.81}] means the accepted values in the first gene start from 1 (inclusive) to 5 (exclusive) while the second one has values between 0.2 (inclusive) and 0.85 (exclusive). For more information, please check the Limit the Gene Value Range section of the documentation.
2. The plot_result() method returns the figure so that the user can save it.
3. Bug fixes in copying elements from the gene space.
4. For a gene with a set of discrete values (more than 1 value) in the gene_space parameter like [0, 1], it was possible that the gene value may not change after mutation. That is if the current value is 0, then the randomly selected value could also be 0. Now, it is verified that the new value is changed. So, if the current value is 0, then the new value after mutation will not be 0 but 1.

Release Date: 20 February 2021

1. 4 new instance attributes are added to hold temporary results after each generation: last_generation_fitness holds the fitness values of the solutions in the last generation, last_generation_parents holds the parents selected from the last generation, last_generation_offspring_crossover holds the offspring generated after applying the crossover in the last generation, and last_generation_offspring_mutation holds the offspring generated after applying the mutation in the last generation. You can access these attributes inside the on_generation() method for example.
2. A bug fixed when the initial_population parameter is used. The bug occurred due to a mismatch between the data type of the array assigned to initial_population and the gene type in the gene_type attribute. Assuming that the array assigned to the initial_population parameter is ((1, 1), (3, 3), (5, 5), (7, 7)) which has type int. When gene_type is set to float, then the genes will not be float but casted to int because the defined array has int type. The bug is fixed by forcing the array assigned to initial_population to have the data type in the gene_type attribute. Check the issue at GitHub: ahmedfgad#27

Thanks to Andrei Rozanski [PhD Bioinformatics Specialist, Department of Tissue Dynamics and Regeneration, Max Planck Institute for Biophysical Chemistry, Germany] for opening my eye to the first change.

Thanks to Marios Giouvanakis, a PhD candidate in Electrical & Computer Engineer, Aristotle University of Thessaloniki (Αριστοτέλειο Πανεπιστήμιο Θεσσαλονίκης), Greece, for emailing me about the second issue.

Release Date: 12 March 2021

1. A new bool parameter called allow_duplicate_genes is supported. If True, which is the default, then a solution/chromosome may have duplicate gene values. If False, then each gene will have a unique value in its solution. Check the Prevent Duplicates in Gene Values section for more details.
2. The last_generation_fitness is updated at the end of each generation not at the beginning. This keeps the fitness values of the most up-to-date population assigned to the last_generation_fitness parameter.

PyGAD 2.14.0 has an issue that is solved in PyGAD 2.14.1. Please consider using 2.14.1 not 2.14.0.

Release Date: 19 May 2021

1. Issue #40 is solved. Now, the None value works with the crossover_type and mutation_type parameters: ahmedfgad#40
2. The gene_type parameter supports accepting a list/tuple/numpy.ndarray of numeric data types for the genes. This helps to control the data type of each individual gene. Previously, the gene_type can be assigned only to a single data type that is applied for all genes. For more information, check the More about the ``gene_type` Parameter <https://pygad.readthedocs.io/en/latest/README_pygad_ReadTheDocs.html#more-about-the-gene-type-parameter>`__ section. Thanks to Rainer Engel for asking about this feature in this discussion: ahmedfgad#43
3. A new bool attribute named gene_type_single is added to the pygad.GA class. It is True when there is a single data type assigned to the gene_type parameter. When the gene_type parameter is assigned a list/tuple/numpy.ndarray, then gene_type_single is set to False.
4. The mutation_by_replacement flag now has no effect if gene_space exists except for the genes with None values. For example, for gene_space=[None, [5, 6]] the mutation_by_replacement flag affects only the first gene which has None for its value space.
5. When an element has a value of None in the gene_space parameter (e.g. gene_space=[None, [5, 6]]), then its value will be randomly generated for each solution rather than being generate once for all solutions. Previously, the gene with None value in gene_space is the same across all solutions
6. Some changes in the documentation according to issue #32: ahmedfgad#32

Release Date: 27 May 2021

1. Some bug fixes when the gene_type parameter is nested. Thanks to Rainer Engel for opening a discussion to report this bug: ahmedfgad#43 (reply in thread)

Rainer Engel helped a lot in suggesting new features and suggesting enhancements in 2.14.0 to 2.14.2 releases.

Release Date: 6 June 2021

1. Some bug fixes when setting the save_best_solutions parameter to True. Previously, the best solution for generation i was added into the best_solutions attribute at generation i+1. Now, the best_solutions attribute is updated by each best solution at its exact generation.

Release Date: 17 June 2021

1. Control the precision of all genes/individual genes. Thanks to Rainer for asking about this feature: ahmedfgad#43 (comment)
2. A new attribute named last_generation_parents_indices holds the indices of the selected parents in the last generation.
3. In adaptive mutation, no need to recalculate the fitness values of the parents selected in the last generation as these values can be returned based on the last_generation_fitness and last_generation_parents_indices attributes. This speeds-up the adaptive mutation.
4. When a sublist has a value of None in the gene_space parameter (e.g. gene_space=[[1, 2, 3], [5, 6, None]]), then its value will be randomly generated for each solution rather than being generated once for all solutions. Previously, a value of None in a sublist of the gene_space parameter was identical across all solutions.
5. The dictionary assigned to the gene_space parameter itself or one of its elements has a new key called "step" to specify the step of moving from the start to the end of the range specified by the 2 existing keys "low" and "high". An example is {"low": 0, "high": 30, "step": 2} to have only even values for the gene(s) starting from 0 to 30. For more information, check the More about the ``gene_space` Parameter <https://pygad.readthedocs.io/en/latest/README_pygad_ReadTheDocs.html#more-about-the-gene-space-parameter>`__ section. ahmedfgad#48
6. A new function called predict() is added in both the pygad.kerasga and pygad.torchga modules to make predictions. This makes it easier than using custom code each time a prediction is to be made.
7. A new parameter called stop_criteria allows the user to specify one or more stop criteria to stop the evolution based on some conditions. Each criterion is passed as str which has a stop word. The current 2 supported words are reach and saturate. reach stops the run() method if the fitness value is equal to or greater than a given fitness value. An example for reach is "reach_40" which stops the evolution if the fitness is >= 40. saturate means stop the evolution if the fitness saturates for a given number of consecutive generations. An example for saturate is "saturate_7" which means stop the run() method if the fitness does not change for 7 consecutive generations. Thanks to Rainer for asking about this feature: ahmedfgad#44
8. A new bool parameter, defaults to False, named save_solutions is added to the constructor of the pygad.GA class. If True, then all solutions in each generation are appended into an attribute called solutions which is NumPy array.
9. The plot_result() method is renamed to plot_fitness(). The users should migrate to the new name as the old name will be removed in the future.
10. Four new optional parameters are added to the plot_fitness() function in the pygad.GA class which are font_size=14, save_dir=None, color="#3870FF", and plot_type="plot". Use font_size to change the font of the plot title and labels. save_dir accepts the directory to which the figure is saved. It defaults to None which means do not save the figure. color changes the color of the plot. plot_type changes the plot type which can be either "plot" (default), "scatter", or "bar". ahmedfgad#47
11. The default value of the title parameter in the plot_fitness() method is "PyGAD - Generation vs. Fitness" rather than "PyGAD - Iteration vs. Fitness".
12. A new method named plot_new_solution_rate() creates, shows, and returns a figure showing the rate of new/unique solutions explored in each generation. It accepts the same parameters as in the plot_fitness() method. This method only works when save_solutions=True in the pygad.GA class's constructor.
13. A new method named plot_genes() creates, shows, and returns a figure to show how each gene changes per each generation. It accepts similar parameters like the plot_fitness() method in addition to the graph_type, fill_color, and solutions parameters. The graph_type parameter can be either "plot" (default), "boxplot", or "histogram". fill_color accepts the fill color which works when graph_type is either "boxplot" or "histogram". solutions can be either "all" or "best" to decide whether all solutions or only best solutions are used.
14. The gene_type parameter now supports controlling the precision of float data types. For a gene, rather than assigning just the data type like float, assign a list/tuple/numpy.ndarray with 2 elements where the first one is the type and the second one is the precision. For example, [float, 2] forces a gene with a value like 0.1234 to be 0.12. For more information, check the More about the ``gene_type` Parameter <https://pygad.readthedocs.io/en/latest/README_pygad_ReadTheDocs.html#more-about-the-gene-type-parameter>`__ section.

The PyGAD library is available at PyPI at this page https://pypi.org/project/pygad. PyGAD is built out of a number of open-source GitHub projects. A brief note about these projects is given in the next subsections.

GitHub Link: https://github.com/ahmedfgad/GeneticAlgorithmPython

GeneticAlgorithmPython is the first project which is an open-source Python 3 project for implementing the genetic algorithm based on NumPy.

GitHub Link: https://github.com/ahmedfgad/NumPyANN

NumPyANN builds artificial neural networks in Python 3 using NumPy from scratch. The purpose of this project is to only implement the forward pass of a neural network without using a training algorithm. Currently, it only supports classification and later regression will be also supported. Moreover, only one class is supported per sample.

GitHub Link: https://github.com/ahmedfgad/NeuralGenetic

NeuralGenetic trains neural networks using the genetic algorithm based on the previous 2 projects GeneticAlgorithmPython and NumPyANN.

GitHub Link: https://github.com/ahmedfgad/NumPyCNN

NumPyCNN builds convolutional neural networks using NumPy. The purpose of this project is to only implement the forward pass of a convolutional neural network without using a training algorithm.

GitHub Link: https://github.com/ahmedfgad/CNNGenetic

CNNGenetic trains convolutional neural networks using the genetic algorithm. It uses the GeneticAlgorithmPython project for building the genetic algorithm.

GitHub Link: https://github.com/ahmedfgad/KerasGA

KerasGA trains Keras models using the genetic algorithm. It uses the GeneticAlgorithmPython project for building the genetic algorithm.

GitHub Link: https://github.com/ahmedfgad/TorchGA

TorchGA trains PyTorch models using the genetic algorithm. It uses the GeneticAlgorithmPython project for building the genetic algorithm.

pygad.torchga: https://github.com/ahmedfgad/TorchGA

If there is an issue using PyGAD, then use any of your preferred option to discuss that issue.

One way is submitting an issue into this GitHub project (github.com/ahmedfgad/GeneticAlgorithmPython) in case something is not working properly or to ask for questions.

If this is not a proper option for you, then check the Contact Us section for more contact details.

PyGAD is actively developed with the goal of building a dynamic library for suporting a wide-range of problems to be optimized using the genetic algorithm.

To ask for a new feature, either submit an issue into this GitHub project (github.com/ahmedfgad/GeneticAlgorithmPython) or send an e-mail to ahmed.f.gad@gmail.com.

Also check the Contact Us section for more contact details.

If you created a project that uses PyGAD, then we can support you by mentioning this project here in PyGAD's documentation.

To do that, please send a message at ahmed.f.gad@gmail.com or check the Contact Us section for more contact details.

Within your message, please send the following details:

• Project title
• Brief description
• Preferably, a link that directs the readers to your project

In this tutorial, we’ll see why mutation with a fixed number of genes is bad, and how to replace it with adaptive mutation. Using the PyGAD Python 3 library, we’ll discuss a few examples that use both random and adaptive mutation.

This tutorial discusses how the genetic algorithm is used to cluster data, starting from random clusters and running until the optimal clusters are found. We'll start by briefly revising the K-means clustering algorithm to point out its weak points, which are later solved by the genetic algorithm. The code examples in this tutorial are implemented in Python using the PyGAD library.

Depending on the nature of the problem being optimized, the genetic algorithm (GA) supports two different gene representations: binary, and decimal. The binary GA has only two values for its genes, which are 0 and 1. This is easier to manage as its gene values are limited compared to the decimal GA, for which we can use different formats like float or integer, and limited or unlimited ranges.

This tutorial discusses how the PyGAD library supports the two GA representations, binary and decimal.

This tutorial introduces PyGAD, an open-source Python library for implementing the genetic algorithm and training machine learning algorithms. PyGAD supports 19 parameters for customizing the genetic algorithm for various applications.

Within this tutorial we'll discuss 5 different applications of the genetic algorithm and build them using PyGAD.

The genetic algorithm (GA) is a biologically-inspired optimization algorithm. It has in recent years gained importance, as it’s simple while also solving complex problems like travel route optimization, training machine learning algorithms, working with single and multi-objective problems, game playing, and more.

Deep neural networks are inspired by the idea of how the biological brain works. It’s a universal function approximator, which is capable of simulating any function, and is now used to solve the most complex problems in machine learning. What’s more, they’re able to work with all types of data (images, audio, video, and text).

Both genetic algorithms (GAs) and neural networks (NNs) are similar, as both are biologically-inspired techniques. This similarity motivates us to create a hybrid of both to see whether a GA can train NNs with high accuracy.

This tutorial uses PyGAD, a Python library that supports building and training NNs using a GA. PyGAD offers both classification and regression NNs.

In this tutorial we'll see how to build a game-playing agent using only the genetic algorithm to play a game called CoinTex, which is developed in the Kivy Python framework. The objective of CoinTex is to collect the randomly distributed coins while avoiding collision with fire and monsters (that move randomly). The source code of CoinTex can be found on GitHub.

The genetic algorithm is the only AI used here; there is no other machine/deep learning model used with it. We'll implement the genetic algorithm using PyGad. This tutorial starts with a quick overview of CoinTex followed by a brief explanation of the genetic algorithm, and how it can be used to create the playing agent. Finally, we'll see how to implement these ideas in Python.

The source code of the genetic algorithm agent is available here, and you can download the code used in this tutorial from here.

PyGAD is an open-source Python library for building the genetic algorithm and training machine learning algorithms. It offers a wide range of parameters to customize the genetic algorithm to work with different types of problems.

PyGAD has its own modules that support building and training neural networks (NNs) and convolutional neural networks (CNNs). Despite these modules working well, they are implemented in Python without any additional optimization measures. This leads to comparatively high computational times for even simple problems.

The latest PyGAD version, 2.8.0 (released on 20 September 2020), supports a new module to train Keras models. Even though Keras is built in Python, it's fast. The reason is that Keras uses TensorFlow as a backend, and TensorFlow is highly optimized.

This tutorial discusses how to train Keras models using PyGAD. The discussion includes building Keras models using either the Sequential Model or the Functional API, building an initial population of Keras model parameters, creating an appropriate fitness function, and more.

PyGAD is a genetic algorithm Python 3 library for solving optimization problems. One of these problems is training machine learning algorithms.

PyGAD has a module called pygad.kerasga. It trains Keras models using the genetic algorithm. On January 3rd, 2021, a new release of PyGAD 2.10.0 brought a new module called pygad.torchga to train PyTorch models. It’s very easy to use, but there are a few tricky steps.

So, in this tutorial, we’ll explore how to use PyGAD to train PyTorch models.

Cómo los algoritmos genéticos pueden competir con el descenso de gradiente y el backprop

Bien que la manière standard d'entraîner les réseaux de neurones soit la descente de gradient et la rétropropagation, il y a d'autres joueurs dans le jeu. L'un d'eux est les algorithmes évolutionnaires, tels que les algorithmes génétiques.

Utiliser un algorithme génétique pour former un réseau de neurones simple pour résoudre le OpenAI CartPole Jeu. Dans cet article, nous allons former un simple réseau de neurones pour résoudre le OpenAI CartPole . J'utiliserai PyTorch et PyGAD .

Cómo los algoritmos genéticos pueden competir con el descenso de gradiente y el backprop

Aunque la forma estandar de entrenar redes neuronales es el descenso de gradiente y la retropropagacion, hay otros jugadores en el juego, uno de ellos son los algoritmos evolutivos, como los algoritmos geneticos.

Usa un algoritmo genetico para entrenar una red neuronal simple para resolver el Juego OpenAI CartPole. En este articulo, entrenaremos una red neuronal simple para resolver el OpenAI CartPole . Usare PyTorch y PyGAD .

파이썬에서 genetic algorithm을 사용하는 패키지들을 다 사용해보진 않았지만, 확장성이 있어보이고, 시도할 일이 있어서 살펴봤다.

이 패키지에서 가장 인상 깊었던 것은 neural network에서 hyper parameter 탐색을 gradient descent 방식이 아닌 GA로도 할 수 있다는 것이다.

개인적으로 이 부분이 어느정도 초기치를 잘 잡아줄 수 있는 역할로도 쓸 수 있고, Loss가 gradient descent 하기 어려운 구조에서 대안으로 쓸 수 있을 것으로도 생각된다.

일단 큰 흐름은 다음과 같이 된다.

사실 완전히 흐름이나 각 parameter에 대한 이해는 부족한 상황

This is a translation of an original English tutorial published at Paperspace: How To Train Keras Models Using the Genetic Algorithm with PyGAD

PyGAD, genetik algoritma oluşturmak ve makine öğrenimi algoritmalarını eğitmek için kullanılan açık kaynaklı bir Python kitaplığıdır. Genetik algoritmayı farklı problem türleri ile çalışacak şekilde özelleştirmek için çok çeşitli parametreler sunar.

PyGAD, sinir ağları (NN’ler) ve evrişimli sinir ağları (CNN’ler) oluşturmayı ve eğitmeyi destekleyen kendi modüllerine sahiptir. Bu modüllerin iyi çalışmasına rağmen, herhangi bir ek optimizasyon önlemi olmaksızın Python’da uygulanırlar. Bu, basit problemler için bile nispeten yüksek hesaplama sürelerine yol açar.

En son PyGAD sürümü 2.8.0 (20 Eylül 2020'de piyasaya sürüldü), Keras modellerini eğitmek için yeni bir modülü destekliyor. Keras Python’da oluşturulmuş olsa da hızlıdır. Bunun nedeni, Keras’ın arka uç olarak TensorFlow kullanması ve TensorFlow’un oldukça optimize edilmiş olmasıdır.

Bu öğreticide, PyGAD kullanılarak Keras modellerinin nasıl eğitileceği anlatılmaktadır. Tartışma, Sıralı Modeli veya İşlevsel API’yi kullanarak Keras modellerini oluşturmayı, Keras model parametrelerinin ilk popülasyonunu oluşturmayı, uygun bir uygunluk işlevi oluşturmayı ve daha fazlasını içerir.

Hogy kontextusba helyezzem a genetikus algoritmusokat, ismételjük kicsit át, hogy hogyan működik a gradient descent és a backpropagation, ami a neurális hálók tanításának általános módszere. Az erről írt cikkemet itt tudjátok elolvasni.

A hálózatok tenyésztéséhez a PyGAD nevű programkönyvtárat használjuk, így mindenek előtt ezt kell telepítenünk, valamint a Tensorflow-t és a Gym-et, amit Colabban már eleve telepítve kapunk.

Maga a PyGAD egy teljesen általános genetikus algoritmusok futtatására képes rendszer. Ennek a kiterjesztése a KerasGA, ami az általános motor Tensorflow (Keras) neurális hálókon történő futtatását segíti. A 47. sorban létrehozott KerasGA objektum ennek a kiterjesztésnek a része és arra szolgál, hogy a paraméterként átadott modellből a második paraméterben megadott számosságú populációt hozzon létre. Mivel a hálózatunk 386 állítható paraméterrel rendelkezik, ezért a DNS-ünk itt 386 elemből fog állni. A populáció mérete 10 egyed, így a kezdő populációnk egy 10x386 elemű mátrix lesz. Ezt adjuk át az 51. sorban az initial_population paraméterben.

PyGAD — это библиотека для имплементации генетического алгоритма. Кроме того, библиотека предоставляет доступ к оптимизированным реализациям алгоритмов машинного обучения. PyGAD разрабатывали на Python 3.

Библиотека PyGAD поддерживает разные типы скрещивания, мутации и селекции родителя. PyGAD позволяет оптимизировать проблемы с помощью генетического алгоритма через кастомизацию целевой функции.

Кроме генетического алгоритма, библиотека содержит оптимизированные имплементации алгоритмов машинного обучения. На текущий момент PyGAD поддерживает создание и обучение нейросетей для задач классификации.

Библиотека находится в стадии активной разработки. Создатели планируют добавление функционала для решения бинарных задач и имплементации новых алгоритмов.

PyGAD разрабатывали на Python 3.7.3. Зависимости включают в себя NumPy для создания и манипуляции массивами и Matplotlib для визуализации. Один из изкейсов использования инструмента — оптимизация весов, которые удовлетворяют заданной функции.

There are different resources that can be used to get started with the genetic algorithm and building it in Python.

To start with coding the genetic algorithm, you can check the tutorial titled Genetic Algorithm Implementation in Python available at these links:

• LinkedIn
• Towards Data Science
• KDnuggets

This tutorial is prepared based on a previous version of the project but it still a good resource to start with coding the genetic algorithm.

Get started with the genetic algorithm by reading the tutorial titled Introduction to Optimization with Genetic Algorithm which is available at these links:

• LinkedIn
• Towards Data Science
• KDnuggets

Read about building neural networks in Python through the tutorial titled Artificial Neural Network Implementation using NumPy and Classification of the Fruits360 Image Dataset available at these links:

• LinkedIn
• Towards Data Science
• KDnuggets

Read about training neural networks using the genetic algorithm through the tutorial titled Artificial Neural Networks Optimization using Genetic Algorithm with Python available at these links:

• LinkedIn
• Towards Data Science
• KDnuggets

To start with coding the genetic algorithm, you can check the tutorial titled Building Convolutional Neural Network using NumPy from Scratch available at these links:

• LinkedIn
• Towards Data Science
• KDnuggets
• Chinese Translation

This tutorial) is prepared based on a previous version of the project but it still a good resource to start with coding CNNs.

Get started with the genetic algorithm by reading the tutorial titled Derivation of Convolutional Neural Network from Fully Connected Network Step-By-Step which is available at these links:

• LinkedIn
• Towards Data Science
• KDnuggets

You can also check my book cited as Ahmed Fawzy Gad 'Practical Computer Vision Applications Using Deep Learning with CNNs'. Dec. 2018, Apress, 978-1-4842-4167-7 which discusses neural networks, convolutional neural networks, deep learning, genetic algorithm, and more.

Find the book at these links:

• Amazon
• Springer
• Apress
• O'Reilly
• Google Books

• E-mail: ahmed.f.gad@gmail.com
• LinkedIn
• Amazon Author Page
• Heartbeat
• Paperspace
• KDnuggets
• TowardsDataScience
• GitHub

Thank you for using PyGAD :)