pymoo 0.6.1.5.dev0__cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

pymoo/algorithms/moo/kgb.py

@@ -0,0 +1,446 @@
+import numpy as np
+import random
+import json
+from pymoo.algorithms.moo.nsga2 import NSGA2
+from pymoo.util.nds.non_dominated_sorting import NonDominatedSorting
+from pymoo.core.population import Population
+try:
+    from sklearn.naive_bayes import GaussianNB
+except:
+    raise "Please install sklearn for KGB: pip install scikit-learn"
+
+
+def euclidean_distance(a, b):
+    a = np.array(a)
+    b = np.array(b)
+    return np.sqrt(np.sum((a - b) ** 2))
+
+
+class KGB(NSGA2):
+    def __init__(
+        self,
+        perc_detect_change=0.1,
+        perc_diversity=0.3,
+        c_size=13,
+        eps=0.0,
+        ps={},
+        perturb_dev=0.1,
+        save_ps=False,
+        **kwargs,
+    ):
+
+        super().__init__(**kwargs)
+        self.PERTURB_DEV = perturb_dev
+        self.PERC_DIVERSITY = perc_diversity
+        self.PERC_DETECT_CHANGE = perc_detect_change
+        self.EPS = eps
+        self.save_ps = save_ps
+
+        self.C_SIZE = c_size
+        self.ps = ps
+        self.nr_rand_solutions = 50 * self.pop_size
+        self.t = 0
+
+        self.rng = np.random.RandomState(self.seed)
+        random.seed(self.seed)
+
+    def setup(self, problem, **kwargs):
+        """
+        Set up the KGB-DMOEA algorithm.
+        :param problem: The optimization problem instance
+        :param kwargs: Additional keyword arguments
+        :return: The result of the superclass setup method
+        """
+        assert (
+            not problem.has_constraints()
+        ), "KGB-DMOEA only works for unconstrained problems."
+        return super().setup(problem, **kwargs)
+    
+
+    def knowledge_reconstruction_examination(self):
+        """
+        Perform the knowledge reconstruction examination.
+        :return: Tuple containing the useful population, useless population, and cluster centroids
+        """
+        clusters = self.ps  # set historical PS set as clusters
+        Nc = self.C_SIZE  # set final nr of clusters
+        size = len(self.ps)  # set size iteration to length of cluster
+        run_counter = 0  # counter variable to give unique key
+
+        # while there are still clusters to be condensed
+        while size > Nc:
+
+            counter = 0
+            min_distance = None
+            min_distance_index = []
+
+            # get clusters that are closest to each other by calculating the euclidean distance
+            for keys_i in clusters.keys():
+                for keys_j in clusters.keys():
+                    if (
+                        clusters[keys_i]["solutions"]
+                        is not clusters[keys_j]["solutions"]
+                    ):
+
+                        dst = euclidean_distance(
+                            clusters[keys_i]["centroid"],
+                            clusters[keys_j]["centroid"],
+                        )
+
+                        if min_distance == None:
+                            min_distance = dst
+                            min_distance_index = [keys_i, keys_j]
+                        elif dst < min_distance:
+                            min_distance = dst
+
+                            min_distance_index = [keys_i, keys_j]
+
+                        counter += 1
+
+            # merge closest clusters
+            for solution in clusters[min_distance_index[1]]["solutions"]:
+                clusters[min_distance_index[0]]["solutions"].append(solution)
+
+            # calculate new centroid for merged cluster
+            clusters[min_distance_index[0]][
+                "centroid"
+            ] = self.calculate_cluster_centroid(
+                clusters[min_distance_index[0]]["solutions"]
+            )
+
+            # remove cluster that was merged
+            del clusters[min_distance_index[1]]
+
+            size -= 1
+            run_counter += 1
+
+        c = []  # list of centroids
+        pop_useful = []
+        pop_useless = []
+
+        # get centroids of clusters
+        for key in clusters.keys():
+            c.append(clusters[key]["centroid"])
+
+        # create pymoo population objected to evaluate centroid solutions
+        centroid_pop = Population.new("X", c)
+
+        # evaluate centroids
+        self.evaluator.eval(self.problem, centroid_pop)
+
+        # do non-dominated sorting on centroid solutions
+        ranking = NonDominatedSorting().do(centroid_pop.get("F"), return_rank=True)[-1]
+
+        # add the individuals from the clusters with the best objective values to the useful population the rest is useless :(
+
+        for idx, rank in enumerate(ranking):
+            if rank == 0:
+                for key in clusters.keys():
+                    if centroid_pop[idx].X == clusters[key]["centroid"]:
+                        for cluster_individual in clusters[key]["solutions"]:
+                            pop_useful.append(cluster_individual)
+            else:
+                for key in clusters.keys():
+                    if centroid_pop[idx].X == clusters[key]["centroid"]:
+                        for cluster_individual in clusters[key]["solutions"]:
+                            pop_useless.append(cluster_individual)
+
+        # return useful and useless population and the centroid solutions
+        return pop_useful, pop_useless, c
+    
+
+    def naive_bayesian_classifier(self, pop_useful, pop_useless):
+        """
+        Train a naive Bayesian classifier using the useful and useless populations.
+        :param pop_useful: Useful population
+        :param pop_useless: Useless population
+        :return: Trained GaussianNB classifier
+        """
+        labeled_useful_solutions = []
+        labeled_useless_solutions = []
+
+        # add labels to solutions
+        for individual in pop_useful:
+            labeled_useful_solutions.append((individual, +1))
+
+        for individual in pop_useless:
+            labeled_useless_solutions.append((individual, -1))
+
+        x_train = []
+        y_train = []
+
+        for i in range(len(labeled_useful_solutions)):
+            x_train.append(labeled_useful_solutions[i][0])
+            y_train.append(labeled_useful_solutions[i][1])
+
+        for i in range(len(labeled_useless_solutions)):
+            x_train.append(labeled_useless_solutions[i][0])
+            y_train.append(labeled_useless_solutions[i][1])
+
+        x_train = np.asarray(x_train)
+        y_train = np.asarray(y_train)
+
+        # fit the naive bayesian classifier with the training data
+        model = GaussianNB()
+        model.fit(x_train, y_train)
+
+        return model
+
+    def add_to_ps(self):
+        """
+        Add the current Pareto optimal set (POS) to the Pareto set (PS) with individual keys.
+        """
+
+        PS_counter = 0
+
+        for individual in self.opt:
+
+            if isinstance(individual.X, list):
+                individual.X = np.asarray(individual.X)
+
+            centroid = self.calculate_cluster_centroid(individual.X)
+
+            self.ps[str(PS_counter) + "-" + str(self.t)] = {
+                "solutions": [individual.X.tolist()],
+                "centroid": centroid,
+            }
+
+            PS_counter += 1
+
+    def predicted_population(self, X_test, Y_test):
+        """
+        Create a predicted population from the test set with positive labels.
+        :param X_test: Test set of features
+        :param Y_test: Test set of labels
+        :return: Predicted population
+        """
+        predicted_pop = []
+        for i in range(len(Y_test)):
+            if Y_test[i] == 1:
+                predicted_pop.append(X_test[i])
+        return predicted_pop
+
+    def calculate_cluster_centroid(self, solution_cluster):
+        """
+        Calculate the centroid for a given cluster of solutions.
+        :param solution_cluster: List of solutions in the cluster
+        :return: Cluster centroid
+        """
+        # Get number of variable shape
+        try:
+            n_vars = len(solution_cluster[0])
+        except TypeError:
+            solution_cluster = np.array(solution_cluster)
+            return solution_cluster.tolist()
+
+        # TODO: this is lazy garbage fix whats coming in
+        cluster = []
+        for i in range(len(solution_cluster)):
+            # cluster.append(solution_cluster[i].tolist())
+            cluster.append(solution_cluster[i])
+        solution_cluster = np.asarray(cluster)
+
+        # Get number of solutions
+        length = solution_cluster.shape[0]
+
+        centroid_points = []
+
+        # calculate centroid for each variable, by taking mean of every variable of cluster
+        for i in range(n_vars):
+            # calculate sum over cluster
+            centroid_points.append(np.sum(solution_cluster[:, i]))
+
+        return [x / length for x in centroid_points]
+
+    def check_boundaries(self, pop):
+        """
+        Check and fix the boundaries of the given population.
+        :param pop: Population to check and fix boundaries
+        :return: Population with corrected boundaries
+        """
+        # check whether numpy array or pymoo population is given
+        if isinstance(pop, Population):
+            pop = pop.get("X")
+
+        # check if any solution is outside the bounds
+        for individual in pop:
+            for i in range(len(individual)):
+                if individual[i] > self.problem.xu[i]:
+                    individual[i] = self.problem.xu[i]
+                elif individual[i] < self.problem.xl[i]:
+                    individual[i] = self.problem.xl[i]
+        return pop
+
+    def random_strategy(self, N_r):
+        """
+        Generate a random population within the problem boundaries.
+        :param N_r: Number of random solutions to generate
+        :return: Randomly generated population
+        """
+        # generate a random population of size N_r
+        # TODO: Check boundaries
+        random_pop = np.random.random((N_r, self.problem.n_var))
+
+        # check if any solution is outside the bounds
+        for individual in random_pop:
+            for i in range(len(individual)):
+                if individual[i] > self.problem.xu[i]:
+                    individual[i] = self.problem.xu[i]
+                elif individual[i] < self.problem.xl[i]:
+                    individual[i] = self.problem.xl[i]
+
+        return random_pop
+
+    def diversify_population(self, pop):
+        """
+        Introduce diversity in the population by replacing a percentage of individuals.
+        :param pop: Population to diversify
+        :return: Diversified population
+        """
+        # find indices to be replaced (introduce diversity)
+        I = np.where(np.random.random(len(pop)) < self.PERC_DIVERSITY)[0]
+        # replace with randomly sampled individuals
+        pop[I] = self.initialization.sampling(self.problem, len(I))
+        return pop
+
+    def _advance(self, **kwargs):
+        """
+        Advance the optimization algorithm by one iteration.
+        """
+        pop = self.pop
+        X, F = pop.get("X", "F")
+
+        # the number of solutions to sample from the population to detect the change
+        n_samples = int(np.ceil(len(pop) * self.PERC_DETECT_CHANGE))
+
+        # choose randomly some individuals of the current population to test if there was a change
+        I = np.random.choice(np.arange(len(pop)), size=n_samples)
+        samples = self.evaluator.eval(self.problem, Population.new(X=X[I]))
+
+        # calculate the differences between the old and newly evaluated pop
+        delta = ((samples.get("F") - F[I]) ** 2).mean()
+
+        # archive the current POS
+        self.add_to_ps()
+
+        # if there is an average deviation bigger than eps -> we have a change detected
+        change_detected = delta > self.EPS
+
+        if change_detected:
+
+            # increase t counter for unique key of PS
+            self.t += 1
+
+            # conduct knowledge reconstruction examination
+            pop_useful, pop_useless, c = self.knowledge_reconstruction_examination()
+
+            # Train a naive bayesian classifier
+            model = self.naive_bayesian_classifier(pop_useful, pop_useless)
+
+            # generate a lot of random solutions with the dimensions of problem decision space
+            X_test = self.random_strategy(self.nr_rand_solutions)
+
+            # introduce noise to vary previously useful solutions
+            noise = np.random.normal(0, self.PERTURB_DEV, self.problem.n_var)
+            noisy_useful_history = np.asarray(pop_useful) + noise
+
+            # check whether solutions are within bounds
+            noisy_useful_history = self.check_boundaries(noisy_useful_history)
+
+            # add noisy useful history to randomly generated solutions
+            X_test = np.vstack((X_test, noisy_useful_history))
+
+            # predict whether random solutions are useful or useless
+            Y_test = model.predict(X_test)
+
+            # create list of useful predicted solutions
+            predicted_pop = self.predicted_population(X_test, Y_test)
+
+            # ------ POPULATION GENERATION --------
+            # take a random sample from predicted pop and known useful pop
+
+            nr_sampled_pop_useful = 0
+            nr_random_filler_solutions = 0
+
+            if len(predicted_pop) >= self.pop_size - self.C_SIZE:
+                init_pop = []
+                predicted_pop = random.sample(
+                    predicted_pop, self.pop_size - self.C_SIZE
+                )
+
+                # add sampled solutions to init_pop
+                for solution in predicted_pop:
+                    init_pop.append(solution)
+
+                # add cluster centroids to init_pop
+                for solution in c:
+                    init_pop.append(np.asarray(solution))
+
+            else:
+
+                # if not enough predicted solutions are available, add all predicted solutions to init_pop
+                init_pop = []
+
+                for solution in predicted_pop:
+                    init_pop.append(solution)
+
+                # add cluster centroids to init_pop
+                for solution in c:
+                    init_pop.append(np.asarray(solution))
+
+            # if there are still not enough solutions in init_pop randomly sample previously useful solutions directly without noise to init_pop
+            if len(init_pop) < self.pop_size:
+
+                # fill up init_pop with randomly sampled solutions from pop_useful
+                if len(pop_useful) >= self.pop_size - len(init_pop):
+
+                    nr_sampled_pop_useful = self.pop_size - len(init_pop)
+
+                    init_pop = np.vstack(
+                        (
+                            init_pop,
+                            random.sample(pop_useful, self.pop_size - len(init_pop)),
+                        )
+                    )
+                else:
+                    # if not enough solutions are available, add all previously known useful solutions without noise to init_pop
+                    for solution in pop_useful:
+                        init_pop.append(solution)
+
+                    nr_sampled_pop_useful = len(pop_useful)
+
+            # if there are still not enough solutions in init_pop generate random solutions with the dimensions of problem decision space
+            if len(init_pop) < self.pop_size:
+
+                nr_random_filler_solutions = self.pop_size - len(init_pop)
+
+                # fill up with random solutions
+                init_pop = np.vstack(
+                    (init_pop, self.random_strategy(self.pop_size - len(init_pop)))
+                )
+
+            # recreate the current population without being evaluated
+            pop = Population.new(X=init_pop)
+
+            # reevaluate because we know there was a change
+            self.evaluator.eval(self.problem, pop)
+
+            # do a survival to recreate rank and crowding of all individuals
+            pop = self.survival.do(self.problem, pop, n_survive=len(pop))
+
+        # create the offsprings from the current population
+        off = self.mating.do(self.problem, pop, self.n_offsprings, algorithm=self)
+        self.evaluator.eval(self.problem, off)
+
+        # merge the parent population and offsprings
+        pop = Population.merge(pop, off)
+
+        # execute the survival to find the fittest solutions
+        self.pop = self.survival.do(
+            self.problem, pop, n_survive=self.pop_size, algorithm=self
+        )
+
+        # dump self.ps to file
+        if self.save_ps:
+            with open("ps.json", "w") as fp:
+                json.dump(self.ps, fp)

pymoo/algorithms/moo/moead.py

@@ -0,0 +1,183 @@
+import numpy as np
+from scipy.spatial.distance import cdist
+
+from pymoo.algorithms.base.genetic import GeneticAlgorithm
+from pymoo.core.algorithm import LoopwiseAlgorithm
+from pymoo.core.duplicate import NoDuplicateElimination
+from pymoo.core.population import Population
+from pymoo.core.selection import Selection
+from pymoo.core.variable import Real, get
+from pymoo.docs import parse_doc_string
+from pymoo.operators.crossover.sbx import SBX
+from pymoo.operators.mutation.pm import PM
+from pymoo.operators.sampling.rnd import FloatRandomSampling
+from pymoo.util.display.multi import MultiObjectiveOutput
+from pymoo.util.reference_direction import default_ref_dirs
+
+
+class NeighborhoodSelection(Selection):
+
+    def __init__(self, prob=1.0) -> None:
+        super().__init__()
+        self.prob = Real(prob, bounds=(0.0, 1.0))
+
+    def _do(self, problem, pop, n_select, n_parents, neighbors=None, **kwargs):
+        assert n_select == len(neighbors)
+        P = np.full((n_select, n_parents), -1)
+
+        prob = get(self.prob, size=n_select)
+
+        for k in range(n_select):
+            if np.random.random() < prob[k]:
+                P[k] = np.random.choice(neighbors[k], n_parents, replace=False)
+            else:
+                P[k] = np.random.permutation(len(pop))[:n_parents]
+
+        return P
+
+
+# =========================================================================================================
+# Implementation
+# =========================================================================================================
+
+class MOEAD(LoopwiseAlgorithm, GeneticAlgorithm):
+
+    def __init__(self,
+                 ref_dirs=None,
+                 n_neighbors=20,
+                 decomposition=None,
+                 prob_neighbor_mating=0.9,
+                 sampling=FloatRandomSampling(),
+                 crossover=SBX(prob=1.0, eta=20),
+                 mutation=PM(prob_var=None, eta=20),
+                 output=MultiObjectiveOutput(),
+                 **kwargs):
+
+        # reference directions used for MOEAD
+        self.ref_dirs = ref_dirs
+
+        # the decomposition metric used
+        self.decomposition = decomposition
+
+        # the number of neighbors considered during mating
+        self.n_neighbors = n_neighbors
+
+        self.neighbors = None
+
+        self.selection = NeighborhoodSelection(prob=prob_neighbor_mating)
+
+        super().__init__(pop_size=len(ref_dirs),
+                         sampling=sampling,
+                         crossover=crossover,
+                         mutation=mutation,
+                         eliminate_duplicates=NoDuplicateElimination(),
+                         output=output,
+                         advance_after_initialization=False,
+                         **kwargs)
+
+    def _setup(self, problem, **kwargs):
+        assert not problem.has_constraints(), "This implementation of MOEAD does not support any constraints."
+
+        # if no reference directions have been provided get them and override the population size and other settings
+        if self.ref_dirs is None:
+            self.ref_dirs = default_ref_dirs(problem.n_obj)
+        self.pop_size = len(self.ref_dirs)
+
+        # neighbours includes the entry by itself intentionally for the survival method
+        self.neighbors = np.argsort(cdist(self.ref_dirs, self.ref_dirs), axis=1, kind='quicksort')[:, :self.n_neighbors]
+
+        # if the decomposition is not set yet, set the default
+        if self.decomposition is None:
+            self.decomposition = default_decomp(problem)
+
+    def _initialize_advance(self, infills=None, **kwargs):
+        super()._initialize_advance(infills, **kwargs)
+        self.ideal = np.min(self.pop.get("F"), axis=0)
+
+    def _next(self):
+        pop = self.pop
+
+        # iterate for each member of the population in random order
+        for k in np.random.permutation(len(pop)):
+            # get the parents using the neighborhood selection
+            P = self.selection.do(self.problem, pop, 1, self.mating.crossover.n_parents, neighbors=[self.neighbors[k]])
+
+            # perform a mating using the default operators - if more than one offspring just pick the first
+            off = np.random.choice(self.mating.do(self.problem, pop, 1, parents=P, n_max_iterations=1))
+
+            # evaluate the offspring
+            off = yield off
+
+            # update the ideal point
+            self.ideal = np.min(np.vstack([self.ideal, off.F]), axis=0)
+
+            # now actually do the replacement of the individual is better
+            self._replace(k, off)
+
+    def _replace(self, k, off):
+        pop = self.pop
+
+        # calculate the decomposed values for each neighbor
+        N = self.neighbors[k]
+        FV = self.decomposition.do(pop[N].get("F"), weights=self.ref_dirs[N, :], ideal_point=self.ideal)
+        off_FV = self.decomposition.do(off.F[None, :], weights=self.ref_dirs[N, :], ideal_point=self.ideal)
+
+        # this makes the algorithm to support constraints - not originally proposed though and not tested enough
+        # if self.problem.has_constraints():
+        #     CV, off_CV = pop[N].get("CV")[:, 0], np.full(len(off_FV), off.CV)
+        #     fmax = max(FV.max(), off_FV.max())
+        #     FV, off_FV = parameter_less(FV, CV, fmax=fmax), parameter_less(off_FV, off_CV, fmax=fmax)
+
+        # get the absolute index in F where offspring is better than the current F (decomposed space)
+        I = np.where(off_FV < FV)[0]
+        pop[N[I]] = off
+
+
+class ParallelMOEAD(MOEAD):
+
+    def __init__(self, ref_dirs, **kwargs):
+        super().__init__(ref_dirs, **kwargs)
+        self.indices = None
+
+    def _infill(self):
+        pop_size, cross_parents, cross_off = self.pop_size, self.mating.crossover.n_parents, self.mating.crossover.n_offsprings
+
+        # do the mating in a random order
+        indices = np.random.permutation(len(self.pop))[:self.n_offsprings]
+
+        # get the parents using the neighborhood selection
+        P = self.selection.do(self.problem, self.pop, self.n_offsprings, cross_parents,
+                              neighbors=self.neighbors[indices])
+
+        # do not any duplicates elimination - thus this results in exactly pop_size * n_offsprings offsprings
+        off = self.mating.do(self.problem, self.pop, 1e12, n_max_iterations=1, parents=P)
+
+        # select a random offspring from each mating
+        off = Population.create(*[np.random.choice(pool) for pool in np.reshape(off, (self.n_offsprings, -1))])
+
+        # store the indices because of the neighborhood matching in advance
+        self.indices = indices
+
+        return off
+
+    def _advance(self, infills=None, **kwargs):
+        assert len(self.indices) == len(infills), "Number of infills must be equal to the one created beforehand."
+
+        # update the ideal point before starting to replace
+        self.ideal = np.min(np.vstack([self.ideal, infills.get("F")]), axis=0)
+
+        # now do the replacements as in the loop-wise version
+        for k, off in enumerate(infills):
+            self._replace(self.indices[k], off)
+
+
+def default_decomp(problem):
+    if problem.n_obj <= 2:
+        from pymoo.decomposition.tchebicheff import Tchebicheff
+        return Tchebicheff()
+    else:
+        from pymoo.decomposition.pbi import PBI
+        return PBI()
+
+
+parse_doc_string(MOEAD.__init__)