fbpx

Simple Genetic Algorithm From Scratch in Python

Last Updated on May 26, 2021

The genetic algorithm is a stochastic global optimization algorithm.

It may be one of the most popular and widely known biologically inspired algorithms, along with artificial neural networks.

The algorithm is a type of evolutionary algorithm and performs an optimization procedure inspired by the biological theory of evolution by means of natural selection with a binary representation and simple operators based on genetic recombination and genetic mutations.

In this tutorial, you will discover the genetic algorithm optimization algorithm.

After completing this tutorial, you will know:

  • Genetic algorithm is a stochastic optimization algorithm inspired by evolution.
  • How to implement the genetic algorithm from scratch in Python.
  • How to apply the genetic algorithm to a continuous objective function.

Let’s get started.

Simple Genetic Algorithm From Scratch in Python

Simple Genetic Algorithm From Scratch in Python
Photo by Magharebia, some rights reserved.

Tutorial Overview

This tutorial is divided into four parts; they are:

  1. Genetic Algorithm
  2. Genetic Algorithm From Scratch
  3. Genetic Algorithm for OneMax
  4. Genetic Algorithm for Continuous Function Optimization

Genetic Algorithm

The Genetic Algorithm is a stochastic global search optimization algorithm.

It is inspired by the biological theory of evolution by means of natural selection. Specifically, the new synthesis that combines an understanding of genetics with the theory.

Genetic algorithms (algorithm 9.4) borrow inspiration from biological evolution, where fitter individuals are more likely to pass on their genes to the next generation.

— Page 148, Algorithms for Optimization, 2019.

The algorithm uses analogs of a genetic representation (bitstrings), fitness (function evaluations), genetic recombination (crossover of bitstrings), and mutation (flipping bits).

The algorithm works by first creating a population of a fixed size of random bitstrings. The main loop of the algorithm is repeated for a fixed number of iterations or until no further improvement is seen in the best solution over a given number of iterations.

One iteration of the algorithm is like an evolutionary generation.

First, the population of bitstrings (candidate solutions) are evaluated using the objective function. The objective function evaluation for each candidate solution is taken as the fitness of the solution, which may be minimized or maximized.

Then, parents are selected based on their fitness. A given candidate solution may be used as parent zero or more times. A simple and effective approach to selection involves drawing k candidates from the population randomly and selecting the member from the group with the best fitness. This is called tournament selection where k is a hyperparameter and set to a value such as 3. This simple approach simulates a more costly fitness-proportionate selection scheme.

In tournament selection, each parent is the fittest out of k randomly chosen chromosomes of the population

— Page 151, Algorithms for Optimization, 2019.

Parents are used as the basis for generating the next generation of candidate points and one parent for each position in the population is required.

Parents are then taken in pairs and used to create two children. Recombination is performed using a crossover operator. This involves selecting a random split point on the bit string, then creating a child with the bits up to the split point from the first parent and from the split point to the end of the string from the second parent. This process is then inverted for the second child.

For example the two parents:

  • parent1 = 00000
  • parent2 = 11111

May result in two cross-over children:

  • child1 = 00011
  • child2 = 11100

This is called one point crossover, and there are many other variations of the operator.

Crossover is applied probabilistically for each pair of parents, meaning that in some cases, copies of the parents are taken as the children instead of the recombination operator. Crossover is controlled by a hyperparameter set to a large value, such as 80 percent or 90 percent.

Crossover is the Genetic Algorithm’s distinguishing feature. It involves mixing and matching parts of two parents to form children. How you do that mixing and matching depends on the representation of the individuals.

— Page 36, Essentials of Metaheuristics, 2011.

Mutation involves flipping bits in created children candidate solutions. Typically, the mutation rate is set to 1/L, where L is the length of the bitstring.

Each bit in a binary-valued chromosome typically has a small probability of being flipped. For a chromosome with m bits, this mutation rate is typically set to 1/m, yielding an average of one mutation per child chromosome.

— Page 155, Algorithms for Optimization, 2019.

For example, if a problem used a bitstring with 20 bits, then a good default mutation rate would be (1/20) = 0.05 or a probability of 5 percent.

This defines the simple genetic algorithm procedure. It is a large field of study, and there are many extensions to the algorithm.

Now that we are familiar with the simple genetic algorithm procedure, let’s look at how we might implement it from scratch.

Genetic Algorithm From Scratch

In this section, we will develop an implementation of the genetic algorithm.

The first step is to create a population of random bitstrings. We could use boolean values True and False, string values ‘0’ and ‘1’, or integer values 0 and 1. In this case, we will use integer values.

We can generate an array of integer values in a range using the randint() function, and we can specify the range as values starting at 0 and less than 2, e.g. 0 or 1. We will also represent a candidate solution as a list instead of a NumPy array to keep things simple.

An initial population of random bitstring can be created as follows, where “n_pop” is a hyperparameter that controls the population size and “n_bits” is a hyperparameter that defines the number of bits in a single candidate solution:

Next, we can enumerate over a fixed number of algorithm iterations, in this case, controlled by a hyperparameter named “n_iter“.

The first step in the algorithm iteration is to evaluate all candidate solutions.

We will use a function named objective() as a generic objective function and call it to get a fitness score, which we will minimize.

We can then select parents that will be used to create children.

The tournament selection procedure can be implemented as a function that takes the population and returns one selected parent. The k value is fixed at 3 with a default argument, but you can experiment with different values if you like.

We can then call this function one time for each position in the population to create a list of parents.

We can then create the next generation.

This first requires a function to perform crossover. This function will take two parents and the crossover rate. The crossover rate is a hyperparameter that determines whether crossover is performed or not, and if not, the parents are copied into the next generation. It is a probability and typically has a large value close to 1.0.

The crossover() function below implements crossover using a draw of a random number in the range [0,1] to determine if crossover is performed, then selecting a valid split point if crossover is to be performed.

We also need a function to perform mutation.

This procedure simply flips bits with a low probability controlled by the “r_mut” hyperparameter.

We can then loop over the list of parents and create a list of children to be used as the next generation, calling the crossover and mutation functions as needed.

We can tie all of this together into a function named genetic_algorithm() that takes the name of the objective function and the hyperparameters of the search, and returns the best solution found during the search.

Now that we have developed an implementation of the genetic algorithm, let’s explore how we might apply it to an objective function.

Genetic Algorithm for OneMax

In this section, we will apply the genetic algorithm to a binary string-based optimization problem.

The problem is called OneMax and evaluates a binary string based on the number of 1s in the string. For example, a bitstring with a length of 20 bits will have a score of 20 for a string of all 1s.

Given we have implemented the genetic algorithm to minimize the objective function, we can add a negative sign to this evaluation so that large positive values become large negative values.

The onemax() function below implements this and takes a bitstring of integer values as input and returns the negative sum of the values.

Next, we can configure the search.

The search will run for 100 iterations and we will use 20 bits in our candidate solutions, meaning the optimal fitness will be -20.0.

The population size will be 100, and we will use a crossover rate of 90 percent and a mutation rate of 5 percent. This configuration was chosen after a little trial and error.

The search can then be called and the best result reported.

Tying this together, the complete example of applying the genetic algorithm to the OneMax objective function is listed below.

Running the example will report the best result as it is found along the way, then the final best solution at the end of the search, which we would expect to be the optimal solution.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

In this case, we can see that the search found the optimal solution after about eight generations.

Genetic Algorithm for Continuous Function Optimization

Optimizing the OneMax function is not very interesting; we are more likely to want to optimize a continuous function.

For example, we can define the x^2 minimization function that takes input variables and has an optima at  f(0, 0) = 0.0.

We can minimize this function with a genetic algorithm.

First, we must define the bounds of each input variable.

We will take the “n_bits” hyperparameter as a number of bits per input variable to the objective function and set it to 16 bits.

This means our actual bit string will have (16 * 2) = 32 bits, given the two input variables.

We must update our mutation rate accordingly.

Next, we need to ensure that the initial population creates random bitstrings that are large enough.

Finally, we need to decode the bitstrings to numbers prior to evaluating each with the objective function.

We can achieve this by first decoding each substring to an integer, then scaling the integer to the desired range. This will give a vector of values in the range that can then be provided to the objective function for evaluation.

The decode() function below implements this, taking the bounds of the function, the number of bits per variable, and a bitstring as input and returns a list of decoded real values.

We can then call this at the beginning of the algorithm loop to decode the population, then evaluate the decoded version of the population.

Tying this together, the complete example of the genetic algorithm for continuous function optimization is listed below.

Running the example reports the best decoded results along the way and the best decoded solution at the end of the run.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

In this case, we can see that the algorithm discovers an input very close to f(0.0, 0.0) = 0.0.

Further Reading

This section provides more resources on the topic if you are looking to go deeper.

Books

API

Articles

Summary

In this tutorial, you discovered the genetic algorithm optimization.

Specifically, you learned:

  • Genetic algorithm is a stochastic optimization algorithm inspired by evolution.
  • How to implement the genetic algorithm from scratch in Python.
  • How to apply the genetic algorithm to a continuous objective function.

Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.

61 Responses to Simple Genetic Algorithm From Scratch in Python

  1. Satish Chhatpar March 4, 2021 at 2:33 am #

    I did not understand above algorithm. Its complex

  2. Wilfredo Yeguez March 4, 2021 at 9:06 am #

    Thanks Jason! You gave me a good push.

  3. Ankita March 5, 2021 at 4:23 am #

    Thankyou so much. Very helpful content for me as i am doing Ph.D in Genetic Algorithm. Could you please help me more. I need some help in further implementation. Mail me as soon as possible.

    • Jason Brownlee March 5, 2021 at 5:35 am #

      I don’t have the capacity to help you with your research project, sorry.

  4. Peter March 5, 2021 at 7:41 am #

    Awesome article, quite large though excellent example to learn from. Thank you

  5. John Lee March 5, 2021 at 12:58 pm #

    Awesome lesson. Thanks!

  6. huibin fu March 5, 2021 at 6:31 pm #

    can i copy the code to my Python? because I want to practice it

  7. Mojtaba March 5, 2021 at 9:46 pm #

    Hi dear Jason.
    Thanks for this helpful tutorial.
    May you give a tutorial on feature selection using genetic algorithms?

  8. Paul Winter March 13, 2021 at 8:03 am #

    Hi Jason, thanks for the great tutorial. I enjoyed reading and typing the code is step by step to really follow along and understand it.

    I modified my genetic_algorithm to also have a decode and bounds input parameter to be able to reuse for both examples. I added a decode for oneup that just teturns the input value, and changed your decode so that the bitstring can be decoded into multiple params of same no of bits.

    Hopefully thats the right direction for reuse.

    I did spot a bug in decode. largest sholuld be (2**n_bits) -1

  9. Junaid Zaheer March 18, 2021 at 4:18 am #

    Hi dear very much difficult to understand such an important topic like genetic algorithm..Any how lots of thank yous to have some light on it..

  10. Arnav Das March 24, 2021 at 6:03 am #

    super cool article jason sir, and really really appreciate for putting everything in code, will help us all a lot in experimenting here and there.

    was just wondering something about these algorithms, would be it fair to say as loss functions are to gradient descent do objective functions also serve the same purpose for genetic algorithms ?

    And compared to genetic algorithms aren’t gradient descent algorithm more objective based, I mean they are solely guided to find the best spot to stop while genetic algorithms more or less rely more on mutations and crossover to reach the same.

    • Jason Brownlee March 25, 2021 at 4:34 am #

      Thanks!

      Yes, sure. Loss function is an objective function for the gradient descent optimizaiton algorithm.

      No, they are just different algorithms. GD uses more information (e.g. derivatives) whereas GAs do not.

  11. JG March 31, 2021 at 6:03 am #

    Hi Jason,

    A great code and introduction to Genetic Algorithms (GA), as a beautiful alternative to Artificial Neural Networks (ANN). Congrats for this post!.

    I am pleasantly surprised about how GA get quick convergence to the minimum quadratic function !.

    In my opinion the main differences between ANN vs GA are:

    with ANN we “map” output vs input with a dataset and a neural model that learn the weights vs GA which solving a “min/max” optimum problem, via “Artificial Gene Selection”. That is, coding “genes” problems in bits > initiating a population > selecting parents via an objective function that evaluated better adaptation to the target > performing Crossover genes > mutation genes > replacing parent population for children population every generation.
    So the key issue is coding the problems variables in bits, to be able to apply crossover and bits mutation methods, plus selecting parents via the better objetive performance.

    – I intuit some “probabilistic” convergence pillars supporting this “Artificial Selection” (or GA) for optimum issues solving vs some SGD and backpropagation methodology (minimum error) as pillars supporting ANN.

    I experiment with other objetive functions such cubic functions, etc. and in all of them the code performing pretty well founded the minimum value very quickly.

    My only concern in terms of “artificial selection” methodology is, of course at least one individual member of the population, get very quickly the minimum searched, but the rest of population (even changing mutation and crossover rate) remain outside this optimum “gene” value, even if I play with different population number, number of generations, mutation and crossover rates, etc…

    so finally we are not able to evolve completely the old population into a “new specie” population, at leat with this chunk of algorithm, but nature can evolve naturally producing new species from old ones :-))

    Thank you for inspiring all of this beautiful issues!
    Regards

    • Jason Brownlee March 31, 2021 at 6:09 am #

      Thanks!

      Yes, I like to think of it as two techniques for solving very different problem types: “function approximation” vs “function optimization”.

      Be careful, tuning GAs can be addictive 🙂

  12. Yessense April 2, 2021 at 10:16 pm #

    There is an error in 63th string:
    >> best, best_eval = 0, objective(pop[0])
    Should be:

    best, best_eval = 0, objective(decode(bounds, n_bits, pop[0])

    • Jason Brownlee April 3, 2021 at 5:32 am #

      Agreed!

      Thanks, fixed.

    • Chris May 24, 2021 at 9:15 pm #

      …and on line 63 a missing parenthesis at the end.

  13. Libo April 18, 2021 at 2:37 pm #

    Hi Jason, Very nice tutorial like all your other tutorials. I have a question, for each new generation, isn’t that we should keep all parents from last generation, plus the current children generation, sorted them according to the scores, keep the top max or min scores? The reason is because that some of the parents are better than child, therefore we want to keep the top performers? Thanks

    • Jason Brownlee April 19, 2021 at 5:48 am #

      Thanks!

      There are many modifications of the algorithm you can make, including keeping the best parents from the previous generation. This is called elitism.

  14. Mark April 20, 2021 at 7:07 pm #

    Hi Jason, can you please make a similar tutorial about Genetic programming, or you can just tell me where the algorithm will have to change to be a genetic programming algorithm not GA

  15. Mohamed April 27, 2021 at 9:57 am #

    Hi, in line 52 at onmax objective function:
    should it be like:

    if score[i] > best_eval:
    best, best_eval = pop[i], scores[i]

    • Jason Brownlee April 28, 2021 at 5:57 am #

      We have inverted the one max objective function to make it minimizing.

  16. Oliver May 1, 2021 at 10:25 pm #

    Thanks jason, I’ve been a long time reader here and I think I’m using your textbook on GA for a class project?

  17. john May 4, 2021 at 12:37 pm #

    hello,, i have syntax error here why?

    for gen in range(n_iter):
    ^
    SyntaxError: invalid syntax

  18. Hamada May 7, 2021 at 8:55 am #

    What is the basis for selecting the values of cross_over and mutation rates ?

    • Jason Brownlee May 8, 2021 at 6:28 am #

      Trial and error, or using values that have historically worked well on other problems.

  19. john May 8, 2021 at 11:10 pm #

    what is the rastrigins function in python?
    how i cam implement it in python

  20. agelos May 9, 2021 at 2:31 am #

    File “”, line 65
    for gen in range(n_iter):
    ^
    SyntaxError: invalid syntax

    in code for continuous function simply copied pasted this error comes up jason

  21. Guilherme May 16, 2021 at 5:39 am #

    Very nice article!

    I’m starting to read about Neural Networks and stumbled upon this page while searching for Genetic Algorithms on Google. It helped me understand some basic concepts.

    Thank you!

  22. Muruganandan S May 17, 2021 at 3:31 am #

    Dear Jason

    Your article is very nice. But, I am not able to go line by line understanding as I am new to the GA. But I got some useful inputs to my work related to stock price predictions. However, I have lots of doubts regrading the implementation of GA in price predictions. Can you help me in this area.

  23. RAHEEL SHAIKH May 25, 2021 at 3:09 pm #

    There is a bug in the code.

    best, best_eval = 0, objective(decode(bounds, n_bits, pop[0])

    should be

    best, best_eval = 0, objective(decode(bounds, n_bits, pop[0]))
    i.e. with last bracket. That is why many people having syntax error.

    Thanks

  24. Yara May 29, 2021 at 10:23 am #

    Hi! First of all, thanks for the tutorial. I’m currently working on an adaptation for a function that depends on 4 variables and having trouble with the decoding function. Is the following right?

    def decode(bounds, n_bits, bitstring):
    decoded = list()
    largest = 4**n_bits-1
    for i in range(len(bounds)):
    # extract the substring
    start, end = i * n_bits, (i * n_bits)+n_bits
    substring = bitstring[start:end]
    # convert bitstring to a string of chars
    chars = ”.join([str(s) for s in substring])
    # convert string to integer
    integer = int(chars, 4)
    # scale integer to desired range
    value = bounds[i][0] + (integer/largest) * (bounds[i][1] – bounds[i][0])
    # store
    decoded.append(value)
    return decoded

    • Jason Brownlee May 30, 2021 at 5:47 am #

      You’re welcome.

      Sorry, I don’t have the capacity to review/debug your extensions. I hope you can understand.

  25. Guixin Liu June 12, 2021 at 8:03 am #

    This is very clear and instructive. I used to study Matlab codes for GA but feel it very difficult. Now I realized it’s not that the algorithm is hard itself, but that the codes I read before was not well written. Thanks!

  26. Jianhua June 12, 2021 at 3:43 pm #

    Hi Jason! Thank you for making this tutorial. I was wondering if it is possible to plot the convergence for your genetic algorithm? If so, how would you implement it?

    • Jason Brownlee June 13, 2021 at 5:47 am #

      Yes, you could save the best fitness in a list each iteration, then plot the list at the end of the run.

  27. Mariona July 6, 2021 at 4:31 am #

    Hi Jason,
    Thank you for sharing this 🙂 I am trying to apply this for a problem with both integer & continuous variables. Any tips on how to do this? I was thinking, in the decode function, only some of the values should be decoded to continuous, the rest should stay as binary or integer.

    • Jason Brownlee July 6, 2021 at 5:50 am #

      Perhaps first decide all to bits to integers, then covert some integers to floats in the required range.

  28. ali July 14, 2021 at 8:04 am #

    thanks for this title
    i have a question , I have some data from a function Can I predict what the actual function is ? use GP

    • Jason Brownlee July 15, 2021 at 5:22 am #

      You can approximate a function that matches the data. This is the goal of applied machine learning (function approximation).

Leave a Reply