This is the first complete program built with DEAP. It will help new users to overview some of the possibilities in DEAP. The problem is very simple, we search for a 1 filled list individual. This problem is widely used in the evolutionary computation community since it is very simple and it illustrates well the potential of evolutionary algorithms.
Here we use the one max problem to show how simple can be an evolutionary algorithm with DEAP. The first thing to do is to elaborate the structures of the algorithm. It is pretty obvious in this case that an individual that can contain a series of booleans is the most interesting kind of structure available. DEAP does not contain any explicit individual structure since it is simply a container of attributes associated with a fitness. Instead, it provides a convenient method for creating types called the creator.
The creator is a class factory that can build at run-time new classes that inherit from a base classe. It is very useful since an individual can be any type of container from list to n-ary tree. The creator allows build complex new structures convenient for evolutionary computation.
Let see an example of how to use the creator to setup an individual that contains an array of booleans and a maximizing fitness. We will first need to import the deap.base and deap.creator modules.
The creator defines at first a single function create() that is used to create types. The create() function takes at least 2 arguments plus additional optional arguments. The first argument name is the actual name of the type that we want to create. The second argument base is the base classe that the new type created should inherit from. Finally the optional arguments are members to add to the new type, for example a fitness for an individual or speed for a particle.
from deap import base
from deap import creator
creator.create("FitnessMax", base.Fitness, weights=(1.0,))
creator.create("Individual", list, fitness=creator.FitnessMax)
The first line creates a maximizing fitness by replacing, in the base type Fitness, the pure virtual weights attribute by (1.0,) that means to maximize a single objective fitness. The second line creates an Individual class that inherits the properties of list and has a fitness attribute of the type FitnessMax that was just created.
In this last step, two things are of major importance. The first is the comma following the 1.0 in the weights declaration, even when implementing a single objective fitness, the weights (and values) must be iterable. We won’t repeat it enough, in DEAP single objective is a special case of multiobjective. The second important thing is how the just created FitnessMax can be used directly as if it was part of the creator. This is not magic.
A Toolbox can be found in the base module. It is intended to store functions with their arguments. The toolbox contains two methods, register() and unregister() that are used to do the tricks.
toolbox = base.Toolbox()
# Attribute generator
toolbox.register("attr_bool", random.randint, 0, 1)
# Structure initializers
toolbox.register("individual", tools.initRepeat, creator.Individual, toolbox.attr_bool, 100)
toolbox.register("population", tools.initRepeat, list, toolbox.individual)
In this code block we registered a generation function and two initialization functions. The generator toolbox.attr_bool() when called, will draw a random integer between 0 and 1. The two initializers for their part will produce respectively initialized individuals and populations.
Again, looking a little closer shows that their is no magic. The registration of tools in the toolbox only associates an alias to an already existing function and freezes part of its arguments. This allows to call the alias as if the majority of the (or every) arguments have already been given. For example, the attr_bool() generator is made from the randint() that takes two arguments a and b, with a <= n <= b, where n is the returned integer. Here, we fix a = 0 and b = 1.
It is the same thing for the initializers. This time, the initRepeat() is frozen with predefined arguments. In the case of the individual() method, initRepeat() takes 3 arguments, a class that is a container – here the Individual is derived from a list –, a function to fill the container and the number of times the function shall be repeated. When called, the individual() method will thus return an individual initialized with what would be returned by 100 calls to the attr_bool() method. Finally, the population() method uses the same paradigm, but we don’t fix the number of individuals that it should contain.
The evaluation function is pretty simple in this case, we need to count the number of ones in the individual. This is done by the following lines of code.
def evalOneMax(individual):
return sum(individual),
The returned value must be an iterable of length equal to the number of objectives (weights).
There is two way of using operators, the first one, is to simply call the function from the tools module and the second one is to register them with their argument in a toolbox as for the initialization methods. The most convenient way is to register them in the toolbox, because it allows to easily switch between operators if desired. The toolbox method is also used in the algorithms, see the one max short version for an example.
Registering the operators and their default arguments in the toolbox is done as follow.
toolbox.register("evaluate", evalOneMax)
toolbox.register("mate", tools.cxTwoPoints)
toolbox.register("mutate", tools.mutFlipBit, indpb=0.05)
toolbox.register("select", tools.selTournament, tournsize=3)
The evaluation is given the alias evaluate. Having a single argument being the individual to evaluate we don’t need to fix any, the individual will be given later in the algorithm. The mutation, for its part, needs an argument to be fixed (the independent probability of each attribute to be mutated indpb). In the algorithms the mutate() function is called with the signature mutant, = toolbox.mutate(mutant). This is the most convenient way because each mutation takes a different number of arguments, having those arguments fixed in the toolbox leave open most of the possibilities to change the mutation (or crossover, or selection, or evaluation) operator later in your researches.
Before evolving it, we need to instantiate a population. This step is done effortless using the method we registered in the toolbox.
pop = toolbox.population(n=300)
pop will be a list composed of 300 individuals, n is the parameter left open earlier in the toolbox. The next thing to do is to evaluate this brand new population.
# Evaluate the entire population
fitnesses = map(toolbox.evaluate, pop)
for ind, fit in zip(pop, fitnesses):
ind.fitness.values = fit
We first map() the evaluation function to every individual, then assign their respective fitness. Note that the order in fitnesses and population are the same.
The evolution of the population is the last thing to accomplish. Let say that we want to evolve for a fixed number of generation MAXGEN, the evolution will then begin with a simple for statement.
for g in range(MAXGEN):
evolve...
Is that simple enough? Lets continue with more complicated things, selecting, mating and mutating the population. The crossover and mutation operators provided within DEAP usually take respectively 2 and 1 individual(s) on input and return 2 and 1 modified individual(s), they also modify inplace these individuals.
In a simple GA, the first step is to select the next generation.
offspring = [toolbox.clone(ind) for ind in toolbox.select(pop, len(pop))]
This step creates an offspring list that is an exact copy of the selected individuals. The toolbox.clone() method ensure that we don’t own a reference to the individuals but an completely independent instance.
Next, a simple GA would replace the parents by the produced children directly in the population. This is what is done by the following lines of code, where a crossover is applied with probability CXPB and a mutation with probability MUTPB. The del statement simply invalidate the fitness of the modified individuals.
# Apply crossover and mutation on the offspring
for child1, child2 in zip(offspring[::2], offspring[1::2]):
if random.random() < CXPB:
toolbox.mate(child1, child2)
del child1.fitness.values
del child2.fitness.values
for mutant in offspring:
if random.random() < MUTPB:
toolbox.mutate(mutant)
del mutant.fitness.values
The population now needs to be re-evaluated, we then apply the evaluation as seen earlier, but this time only on the individuals with an invalid fitness.
# Evaluate the individuals with an invalid fitness
invalid_ind = [ind for ind in offspring if not ind.fitness.valid]
fitnesses = map(toolbox.evaluate, invalid_ind)
for ind, fit in zip(invalid_ind, fitnesses):
ind.fitness.values = fit
And finally, last but not least, we replace the old population by the offspring.
pop = offspring
This is the end of the evolution part, it will continue until the predefined number of generation are accomplished.
Although, some statistics may be gathered on the population, the following lines print the min, max, mean and standard deviation of the population.
# Gather all the fitnesses in one list and print the stats
fits = [ind.fitness.values[0] for ind in pop]
length = len(pop)
mean = sum(fits) / length
sum2 = sum(x*x for x in fits)
std = abs(sum2 / length - mean**2)**0.5
print " Min %s" % min(fits)
print " Max %s" % max(fits)
print " Avg %s" % mean
print " Std %s" % std
A Statistics object has been defined to facilitate how statistics are gathered. It is not presented here so that we can focus on the core and not the gravitating helper objects of DEAP. The complete One Max Genetic Algorithm code is available. It may be a little different but it does the overall same thing.