genetic

(ver. 0.1.dev2)

This package is intended for numerical optimisation. The main goals are flexibility and ease of use. A while ago I needed a genetic algorithm that would allow arbitrary combinations of objects (parameters) in the genome. The lack of such implementation in Python (or my failure to find one) pushed me towards developing this package. As of now it contains 4 main modules

• individuals

  • BaseIndividual - the base individual class. Any individual you'd wish to create must subclass this base class and implement all its methods and properties to make it compatible with any proper population. Refer to examples and docs for further details
  • SingleChromosomeIndividual – a basic individual that only has one chromosome. It is ageless. Mating is based on the recombination.binomal function.

• populations todo

• recombination todo

• selection todo

• util todo

---

Example 1. Optimising sum

First, let's create an individual. To do that we need an engine that will generate gene values

>>> import random

>>> def engine(_):
        return random.randint(0, 50)

Note that the engine must have a single argument. That is because in some cases one may want to generate a new value based on the current state of the genome. We will generate the values totally randomly.

Let's create the first individual. Let it have 10 genes. We can provide unique engine for each gene, but in this case all genes will have the same engine. And let's set the random mutation rate to 0.1.

>>> from genetic.individuals import SingleChromosomeIndividual


>>> indiv = SingleChromosomeIndividual(engine, 0.1, 10)

Here we have our first individual.

Now, let's move on to the population. We will need to create the target (fitness) function to maximise.

>>> def fitness(x):
        """
        :type x: SingleChromosomeIndividual
        """
        return - abs(200 - sum(x.genome))

This function takes a SingleChromosomeIndividual instance and evaluates the negative of the absolute difference between 200 and the sum of genes (numbers in this case).

Now we need a selection model. Let's pick one from the selection module

>>> from genetic.selection import bimodal

>>> selection = bimodal(fittest_fraction=0.2, other_random_survival=0.05)

Here we used a selection factory, that has two parameters: the fraction of the fittest individuals that survive a generation and the fraction of random survivors in the rest of the population.

Now we can start the population. We need to pass at lest 2 individuals to begin with, so we'll just take a copy of the same one. Let's have 100 individuals in the population. And let's keep track of 10 legends

>>> from genetic.populations import PanmicticPopulation

>>> ancestors = [indiv] * 2
>>> population = PanmicticPopulation(ancestors, 100, fitness, selection, 10)

Now, let's make it evolve for 10 generations

>>> average_fitness = list(population.evolve(10))

Note that the PanmicticPopulation.evolve method returns a lazy generator, hence to make the evolution happen, you need to make it generate values (to iterate over it). Our errors are:

>>> print(average_fitness)

[-59.979999999999997,
 -51.509999999999998,
 -41.18,
 -36.960000000000001,
 -30.359999999999999,
 -28.460000000000001,
 -28.82,
 -27.27,
 -28.5,
 -33.960000000000001]

Let's look at the legends' scores

>>> print([legend[0] for legend in population.legends])

[0, 0, 0, -1, -1, -1, -1, -1, -1, -1]

As you see, we already have 3 optimal solutions. Let's take a look at the first one

>>> print(population.legends[0][1].genome)

(8, 31, 11, 8, 48, 2, 17, 25, 43, 7)

---

Example 2. Optimising neural network architecture

todo