reproduce_multiclass.py

from gene import NodeGene, ConnectionGene
from genome_multiclass import GenomeMultiClass
from species import SpeciesSet
import random
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import math
import copy
import math


class ReproduceMultiClass:
    """
    Will handle reproduction of the genomes
    """

    def __init__(self, stagnation, config):
        self.global_innovation_number = 0
        self.stagnation = stagnation
        self.ancestors = {}
        self.genome_indexer = 0
        self.config = config
        # Key: The tuple of the connection e.g. (1,3) value: the innovation number
        self.innovation_tracker = {}

    def create_new_population(self, population_size, num_features, num_classes):
        population = {}

        source_node_list = []
        output_node_list = []

        # Create the source and output nodes
        for node_id in range(num_features + num_classes):
            if node_id < num_features:
                source_node_list.append(NodeGene(node_id=node_id, node_type='source'))
            else:
                output_node_list.append(NodeGene(node_id=node_id, node_type='output', bias=1))

        # Save innovations on population creation
        for source_node in source_node_list:
            for output_node in output_node_list:
                # Increment for the new innovation
                self.global_innovation_number += 1
                # The output node will always have the node_id equal to the number of features
                self.innovation_tracker[(source_node.node_id, output_node.node_id)] = self.global_innovation_number

        connection_list = []
        # For each feature there will be a connection to the output
        for source_node in source_node_list:
            for output_node in output_node_list:
                connection = (source_node.node_id, output_node.node_id)
                # The connection was already saved, so this should return true
                assert (connection in self.innovation_tracker)
                connection_list.append(ConnectionGene(input_node=source_node.node_id, output_node=output_node.node_id,
                                                      innovation_number=self.innovation_tracker[connection],
                                                      enabled=True))
        all_nodes_list = source_node_list + output_node_list
        # Create a population of size population_size
        for index in range(population_size):
            # Deep copies otherwise changing the connection weight change's it for every genome that has the same
            # reference to the class
            deep_copy_connections = copy.deepcopy(connection_list)
            deep_copy_nodes = copy.deepcopy(all_nodes_list)
            # Set all the connections to a random weight for each genome
            for connection in deep_copy_connections:
                connection.weight = np.random.randn()
            # Increment since the index value has been assigned
            self.genome_indexer += 1

            # Create the genome
            population[index] = GenomeMultiClass(connections=deep_copy_connections, nodes=deep_copy_nodes,
                                                 key=self.genome_indexer)

        self.show_population_weight_distribution(population=population)

        return population

    @staticmethod
    def show_population_weight_distribution(population):
        # See the spread of starting weights
        list_of_weights = []
        for genome in population.values():
            for connection in genome.connections.values():
                list_of_weights.append(connection.weight)

        sns.distplot(list_of_weights)
        plt.title('Weight distribution of connections in each population member')
        plt.xlabel('Connection weight values')
        plt.show()

    @staticmethod
    def compute_adjusted_species_sizes(adjusted_species_fitnesses, previous_species_sizes, population_size,
                                       min_species_size):
        """
        Compute the number of offspring per species, proportional to their fitnesses (See page 110 of NEAT paper)
        :param adjusted_species_fitnesses:
        :param previous_species_sizes:
        :param population_size:
        :param min_species_size:
        :return:
        """

        # Sum all the remaining adjusted species fitnesses
        adjusted_fitness_sum = sum(adjusted_species_fitnesses)

        adjusted_species_sizes = []

        for adjusted_fitness, previous_size in zip(adjusted_species_fitnesses, previous_species_sizes):
            if adjusted_fitness_sum is not None:
                # Calculate the adjusted species size for how much of the overall fitness they account for. If this
                # value is less than the min_species_size then we set it to that instead
                species_size = max(min_species_size, ((adjusted_fitness / adjusted_fitness_sum) * population_size))

            else:
                species_size = min_species_size

            # difference = (species_size - previous_size) * 0.5
            # rounded_difference = int(round(difference))
            # adjusted_size = previous_size
            # if abs(rounded_difference) > 0:
            #     adjusted_size += rounded_difference
            # elif difference > 0:
            #     adjusted_size += 1
            # elif difference < 0:
            #     adjusted_size -= 1
            #
            # adjusted_species_sizes.append(adjusted_size)
            # TODO: This allows for the fitter species to more aggressively have more population to create. If you want this behaviour comment out everything above until the end of the if else statement and uncomment this
            adjusted_species_sizes.append(round(species_size))

        # Normalize the spawn amounts so that the next generation is roughly
        # the population size requested by the user.
        total_adjusted_size = sum(adjusted_species_sizes)
        norm = population_size / total_adjusted_size
        adjusted_species_sizes = [max(min_species_size, int(round(n * norm))) for n in adjusted_species_sizes]

        print('NEW POPULATION SIZE: {}'.format(sum(adjusted_species_sizes)))

        return adjusted_species_sizes

    def get_non_stagnant_species(self, species_set, generation):
        """
        Checks which species are stagnant ant returns the ones which aren't
        :param generation: Which generation number it is
        :param species_set: The species set instance which stores all the species
        :return: A list of non stagnant species
        """
        # Keeps track of all the fitnesses for the genomes in the population
        all_fitnesses = []
        # Keeps track of the species which aren't stagnant
        remaining_species = []

        # (Id, species instance, boolean)
        for species_id, species, is_stagnant in self.stagnation.update(species_set=species_set, generation=generation,
                                                                       config=self.config):
            # if is_stagnant:
            #     # TODO: What to do here??
            #     pass
            # else:
            #     # Save all the fitness in the species that isn't stagnant
            #     all_fitnesses += [member.fitness for member in species.members.values()]
            #     remaining_species.append(species)

            # Only save species if it is not stagnant
            if not is_stagnant:
                # Save all the fitness in the species that isn't stagnant
                all_fitnesses += [member.fitness for member in species.members.values()]
                remaining_species.append(species)

        # The case where there are no species left
        if not remaining_species:
            # TODO: Would this ever come here?
            raise Exception('There are no remaining species in the reproduce function')

        return all_fitnesses, remaining_species

    def get_adjusted_species_sizes(self, all_fitnesses, remaining_species, population_size):
        """
        Adjusts the size of the species for their fitness values
        :param all_fitnesses: A list of all fitness values for all genomes in the population
        :param remaining_species: A list of species which aren't stagnant
        :param population_size: The population size
        :return: A list of sizes for the new remaining species, adjusted for their respective fitness values
        """

        # Find min and max fitness across the entire population. We use this for explicit fitness sharing.
        min_genome_fitness = min(all_fitnesses)
        max_genome_fitness = max(all_fitnesses)

        # TODO: The value 1.0 is arbtrirary from the neat python package from previous. Should let it be configurable?
        fitness_range = max(1.0, max_genome_fitness - min_genome_fitness)

        # TODO: Not sure if this is the right method to do adjusted fitness
        for species in remaining_species:
            # The adjusted fitness is the mean of the species members fitnesses TODO: Is this correct?
            mean_species_fitness = np.mean([member.fitness for member in species.members.values()])
            species.adjusted_fitness = (mean_species_fitness - min_genome_fitness) / fitness_range

        adjusted_species_fitnesses = [species.adjusted_fitness for species in remaining_species]

        # Get a list of the amount of members in each of the remaining species
        previous_species_sizes = [len(species.members) for species in remaining_species]

        # If the sum of the adjusted species fitnesses is less than 0.1, it suggests there isn't much fitness variation
        # in the population. Thus we put an artificial barrier to the min species size because there is no species that
        # entirely beats all other species
        # TODO: 0.1 is an random number and should be configurable
        adjusted_species_sum = sum(adjusted_species_fitnesses)
        if adjusted_species_sum < 0.1:
            min_species_size = 2
        else:
            min_species_size = self.config.min_species_size

        adjusted_species_sizes = self.compute_adjusted_species_sizes(
            adjusted_species_fitnesses=adjusted_species_fitnesses, min_species_size=min_species_size,
            previous_species_sizes=previous_species_sizes, population_size=population_size)

        return adjusted_species_sizes

    def get_new_population(self, adjusted_species_sizes, remaining_species, species_set, generation_tracker,
                           backprop_mutation):
        """
        Creates the dictionary of the new genomes for the next generation population
        :param: genetation_tracker:
        :param adjusted_species_sizes:
        :param remaining_species:
        :param species_set:
        :param new_population:
        :return:
        """
        new_population = {}

        for species_size, species in zip(adjusted_species_sizes, remaining_species):

            # TODO: Uncomment if you removed min_species_size
            # assert (species_size > 0)
            if species_size > 0:

                # List of old species members
                old_species_members = list(species.members.values())
                # Reset the members for the current species
                species.members = {}
                # Save the species in the species set object
                species_set.species[species.key] = species

                # Sort the members into the descending fitness
                old_species_members.sort(reverse=True, key=lambda x: x.fitness)

                # Double check that it is descending
                if len(old_species_members) > 1:
                    assert (old_species_members[0].fitness >= old_species_members[1].fitness)

                # If we have specified a number of genomes to carry over, carry them over to the new population
                num_genomes_without_crossover = int(
                    round(species_size * self.config.chance_for_mutation_without_crossover))
                if num_genomes_without_crossover > 0:

                    for member in old_species_members[:num_genomes_without_crossover]:

                        # Check if we should carry over a member un-mutated or not
                        if not self.config.keep_unmutated_top_percentage:
                            child = copy.deepcopy(member)

                            child.mutate(reproduction_instance=self,
                                         innovation_tracker=self.innovation_tracker, config=self.config,
                                         backprop_mutation=backprop_mutation)

                            if not child.check_connection_enabled_amount() and not child.check_num_paths(
                                    only_add_enabled_connections=True):
                                raise Exception('This child has no enabled connections')

                            new_population[child.key] = child
                            self.ancestors[child.key] = ()
                            # new_population[member.key] = member
                            species_size -= 1
                            assert (species_size >= 0)
                        else:
                            # Else we just add the current member to the new population
                            new_population[member.key] = member
                            species_size -= 1
                            assert (species_size >= 0)

                # If there are no more genomes for the current species, then restart the loop for the next species
                if species_size <= 0:
                    continue

                # Only use the survival threshold fraction to use as parents for the next generation.
                reproduction_cutoff = int(math.ceil((1 - self.config.chance_for_mutation_without_crossover) *
                                                    len(old_species_members)))

                # Need at least two parents no matter what the previous result
                reproduction_cutoff = max(reproduction_cutoff, 2)
                old_species_members = old_species_members[:reproduction_cutoff]

                # Randomly choose parents and choose whilst there can still be additional genomes for the given species
                while species_size > 0:
                    species_size -= 1

                    # TODO: If you don't allow them to mate with themselves then it's a problem because if the species previous
                    # TODO: size is 1, then how can you do with or without crossover?
                    parent_1 = copy.deepcopy(random.choice(old_species_members))
                    parent_2 = copy.deepcopy(random.choice(old_species_members))

                    # Has to be a deep copy otherwise the connections which are crossed over are also modified if mutation
                    # occurs on the child.
                    parent_1 = copy.deepcopy(parent_1)
                    parent_2 = copy.deepcopy(parent_2)

                    self.genome_indexer += 1
                    genome_id = self.genome_indexer

                    child = GenomeMultiClass(key=genome_id)
                    # TODO: Save the parent_1 and parent_2 mutation history as well as what connections they had
                    # Create the genome from the parents
                    num_connections_enabled = child.crossover(genome_1=parent_1, genome_2=parent_2, config=self.config)

                    # If there are no connections enabled we forget about this child and don't add it to the existing
                    # population
                    if num_connections_enabled:
                        child.mutate(reproduction_instance=self,
                                     innovation_tracker=self.innovation_tracker, config=self.config,
                                     generation_tracker=generation_tracker, backprop_mutation=backprop_mutation)

                        if not child.check_connection_enabled_amount() and not child.check_num_paths(
                                only_add_enabled_connections=True):
                            raise Exception('This child has no enabled connections')

                        new_population[child.key] = child
                        self.ancestors[child.key] = (parent_1.key, parent_2.key)
                    else:
                        # Else if the crossover resulted in an invalid genome.
                        assert num_connections_enabled == 0
                        species_size += 1
                        self.genome_indexer -= 1

        return new_population

    def reproduce(self, species_set, population_size, generation, generation_tracker, backprop_mutation=False):
        """
        Handles reproduction of a population
        :param generation_tracker: An class instance which keeps track of certain parameters for each generation
        :param generation: Which generation number it is
        :param species_set: The SpeciesSet instance which keeps track of species
        :param population_size: The population size
        :return: A new population
        """
        # Check it is a class instance
        assert (isinstance(species_set, SpeciesSet))

        all_fitnesses, remaining_species = self.get_non_stagnant_species(species_set=species_set, generation=generation)

        adjusted_species_sizes = self.get_adjusted_species_sizes(all_fitnesses=all_fitnesses,
                                                                 population_size=population_size,
                                                                 remaining_species=remaining_species)

        # Set the species dict to an empty one for now as the new species will be configured later
        species_set.species = {}

        # Keeps track of the new population (key, object)
        new_population = self.get_new_population(adjusted_species_sizes=adjusted_species_sizes, species_set=species_set,
                                                 remaining_species=remaining_species,
                                                 generation_tracker=generation_tracker,
                                                 backprop_mutation=backprop_mutation)

        return new_population