Source code for sigmaepsilon.math.optimize.ga

"""Genetic algorithm base classes and population data structures."""

from typing import Iterable, Callable, Tuple, Generator
from types import NoneType
from numbers import Number
import operator
from enum import Enum, unique
from concurrent.futures import ProcessPoolExecutor

import numpy as np
from numpy import ndarray
from numpy.random import Generator as RNG
from pydantic import BaseModel, Field

from ..function import Function
from .state import OptimizerState
from .selection import SelectionStrategy, TournamentSelection

__all__ = ["GeneticAlgorithm", "Genom"]


def even(n: Number) -> bool:
    return n % 2 == 0


def odd(n: Number) -> bool:
    return not even(n)


[docs] class Genom(BaseModel): """A data class for members of a population.""" phenotype: list[float] = Field(default_factory=list) genotype: list[int | float] = Field(default_factory=list) fitness: float age: int = Field(default=0) index: int = Field(default=-1) def __eq__(self, other) -> bool: """Return whether self and other have the same genotype.""" if not isinstance(other, Genom): return False return np.all(self.genotype == other.genotype) def __hash__(self): """Return a hash of self based on the genotype.""" return hash(tuple(float(x) for x in self.genotype)) def __gt__(self, other): """Return whether self's fitness is greater than other's fitness.""" if not isinstance(other, Genom): raise TypeError( f"This operation is not supported between instances of {type(other)} and {type(self)}." ) return np.all(self.fitness > other.fitness) def __lt__(self, other): """Return whether self's fitness is less than other's fitness.""" if not isinstance(other, Genom): raise TypeError( f"This operation is not supported between instances of {type(other)} and {type(self)}." ) return np.all(self.fitness < other.fitness) def __ge__(self, other): """Return whether self's fitness is greater than or equal to other's fitness.""" if not isinstance(other, Genom): raise TypeError( f"This operation is not supported between instances of {type(other)} and {type(self)}." ) return np.all(self.fitness >= other.fitness) def __le__(self, other): """Return whether self's fitness is less than or equal to other's fitness.""" if not isinstance(other, Genom): raise TypeError( f"This operation is not supported between instances of {type(other)} and {type(self)}." ) return np.all(self.fitness <= other.fitness) @property def fittness(self) -> float: # pragma: no cover """Return the fitness value (deprecated alias for :attr:`fitness`).""" import warnings warnings.warn( "'fittness' was a typo and is deprecated; it will be removed in a future version. Use 'fitness' instead.", DeprecationWarning, stacklevel=2, ) return self.fitness
[docs] class GeneticAlgorithm: """ Base class for Genetic Algorithms (GA). Use this as a base class to your custom implementation of a GA. The class has 3 representation-specific extension points that a subclass must implement to yield a working genetic algorithm: :func:`populate`, :func:`crossover` and :func:`mutate` (their base implementations raise ``NotImplementedError``). :func:`encode`/:func:`decode` default to the identity mapping (suitable for representations where genotype and phenotype coincide, e.g. real-valued encodings) and :func:`select` already comes with a working, pluggable default (see :attr:`selection_strategy` / :class:`~sigmaepsilon.math.optimize.selection.SelectionStrategy`); override them only if you need representation-specific behavior. It is also possible to use a custom stopping criteria by overriding :func:`stopping_criteria`. See the :class:`~sigmaepsilon.math.optimize.bitchromosome.BitChromosomeGeneticAlgorithm` (shared bit-chromosome machinery, with :class:`~sigmaepsilon.math.optimize.bga.BinaryGeneticAlgorithm` and :class:`~sigmaepsilon.math.optimize.iga.IntegerGeneticAlgorithm` as its concrete, continuous/integer subclasses) and :class:`~sigmaepsilon.math.optimize.rga.RealValuedGeneticAlgorithm` classes for examples. .. note:: This class is designed for maximizing the objective function. To minimize it, either negate the objective function or pass ``minimize=True`` when instantiating the class. Parameters ---------- fnc: Callable The function to evaluate. It is assumed, that the function expects and N number of scalar arguments as a 1d iterable. ranges: Iterable Ranges for each scalar argument to the objective function. length: int, Optional Chromosome length. The higher the value, the more precision. Default is 5. p_c: float, Optional Probability of crossover. Default is 1. p_m: float, Optional Probability of mutation. Default is 0.2. nPop: int, Optional The size of the population. Default is 100. maxiter: int, Optional The maximum number of iterations. Default is 200. miniter: int, Optional The minimum number of iterations. Default is 100. elitism: float | int | None, Optional Determines the portion of the population designated as elite, which automatically survives to the next generation. If less than or equal to 1, it specifies a fraction of the population. If greater than 1, it indicates the exact number of individuals to be selected as elite. The default value of 1 assures that the reigning champion is always preserved. To turn this off, det the value to None. Default is 1. ftol: float, Optional Torelance for floating point operations. Default is 1e-12. maxage: int, Optional The age is the maximum number of generations a candidate spends at the top (being the best candidate). Default is 5. minimize: bool, Optional If True, the objective function is minimized. Default is False. seed: int | numpy.random.SeedSequence | numpy.random.Generator | None, Optional A seed for a per-instance :func:`numpy.random.default_rng`, used for every stochastic operation instead of the global :mod:`numpy.random` state. Passing the same seed makes runs reproducible, and lets multiple instances draw independent random streams in the same process. Default is None (nondeterministic). selection_strategy: :class:`~sigmaepsilon.math.optimize.selection.SelectionStrategy`, Optional The strategy used by the default :func:`select` implementation to pick the survivors of a generation. Default is :class:`~sigmaepsilon.math.optimize.selection.TournamentSelection`. vectorized: bool, Optional If True, :func:`evaluate` calls the objective function once with the whole ``(nPop, dim)`` array of phenotypes and expects it to return one fitness value per row, instead of calling it once per individual. Default is False. n_jobs: int, Optional If different from 1 and ``vectorized`` is False, :func:`evaluate` calls the objective function once per individual in parallel worker processes (using :class:`concurrent.futures.ProcessPoolExecutor`); -1 means "use all available CPUs". The objective function must be picklable (e.g. a module-level function, not a lambda or a closure). Default is 1 (sequential evaluation). Note ---- Be cautious what you use a genetic algorithm for. Like all metahauristic methods, a genetic algorithm can be wery demanding on the computational side. If the objective function takes a lot of time to evaluate, it is probably not a good idea to use a heuristic approach, unless you have a dedicated evaluator that is able to run efficiently for a large number of problems or if the long running time is not an issue. If you want to customize the way the objective is evaluated, override the :func:`evaluate` method. See Also -------- :class:`~sigmaepsilon.math.optimize.bga.BinaryGeneticAlgorithm` :class:`~sigmaepsilon.math.optimize.iga.IntegerGeneticAlgorithm` :class:`~sigmaepsilon.math.optimize.rga.RealValuedGeneticAlgorithm` """ @unique class Status(Enum): """Status codes for the genetic algorithm.""" INITIALIZED = 0 MAX_ITERATIONS_REACHED = 1 CONVERGED = 2 ERROR = -1 __slots__ = [ "fnc", "ranges", "dim", "length", "p_c", "p_m", "nPop", "_genotypes", "_phenotypes", "_fitness", "_champion", "_evolver", "maxiter", "miniter", "elitism", "maxage", "_is_symbolic_Function", "_celebrate_op", "_minimize", "_state", "_status", "_rng", "selection_strategy", "vectorized", "n_jobs", ] def __init__( self, fnc: Callable, ranges: Iterable, *, length: int = 5, p_c: float = 1, p_m: float = 0.2, nPop: int = 100, maxiter: int = 200, miniter: int = 0, elitism: int | float | NoneType = 1, maxage: int = 5, minimize: bool = False, seed: int | np.random.SeedSequence | RNG | NoneType = None, selection_strategy: SelectionStrategy | NoneType = None, vectorized: bool = False, n_jobs: int = 1, ): super().__init__() self.fnc = fnc self.ranges = np.array(ranges) self.dim = getattr(fnc, "dimension", self.ranges.shape[0]) self.length = length self._rng = seed if isinstance(seed, RNG) else np.random.default_rng(seed) self.selection_strategy = ( selection_strategy if selection_strategy is not None else TournamentSelection() ) self.vectorized = vectorized self.n_jobs = n_jobs if odd(nPop): nPop += 1 if odd(int(nPop / 2)): nPop += 2 assert nPop % 4 == 0 assert nPop >= 4 self._is_symbolic_Function = isinstance(fnc, Function) and fnc.is_symbolic self.nPop = nPop self.p_c = None self.p_m = None self._genotypes = None self._phenotypes = None self._fitness = None self._champion: Genom | NoneType = None self._celebrate_op = None self._minimize = False self.elitism = None self._state = None self._status = None self.set_solution_params( p_c=p_c, p_m=p_m, maxiter=maxiter, miniter=miniter, elitism=elitism, maxage=maxage, minimize=minimize, ) self.reset() @property def state(self) -> OptimizerState: """Return the state of the optimizer.""" return self._state @property def rng(self) -> RNG: """ Return the random number generator of the instance. All stochastic operations (population initialization, crossover, mutation, selection) must draw from this generator rather than the global :mod:`numpy.random` state, so that runs are reproducible (via the `seed` constructor argument) and independent instances don't interfere with each other's randomness. """ return self._rng @property def diversity(self) -> float: """ Return a simple measure of the phenotypic diversity of the current population. Computed as the mean, over all dimensions, of the per-dimension standard deviation of the phenotypes. A value close to zero indicates a converged, homogeneous population; this can be used, in addition to champion age, as a signal for premature convergence. """ phenotypes = np.asarray(self.phenotypes, dtype=float) if phenotypes.size == 0: return 0.0 return float(np.mean(np.std(phenotypes, axis=0))) @property def nIter(self) -> int: # pragma: no cover """For backwards compatibility. Returns the number of iterations performed.""" return self.state.n_iter @property def champion(self) -> Genom: """Return the genom of the champion.""" return self._champion @property def genotypes(self) -> Iterable: """Return the genotypes of the population.""" return self._genotypes @genotypes.setter def genotypes(self, value: Iterable) -> None: """Set the genotypes of the population.""" self._genotypes = value self._phenotypes = None self._fitness = None @property def phenotypes(self) -> Iterable: """Return the phenotypes of the population.""" if self._phenotypes is None: genotypes = self.genotypes if genotypes is not None: self._phenotypes = self.decode(genotypes) return self._phenotypes @property def fitness(self) -> ndarray: """ Return the actual fitness values of the population. Or the fitness of the population described by the argument `phenotypes`. """ if self._fitness is not None: return self._fitness self._fitness = self.evaluate(self.phenotypes) return self._fitness @property def fittness(self) -> ndarray: # pragma: no cover """Return the fitness values (deprecated alias for :attr:`fitness`).""" import warnings warnings.warn( "'fittness' was a typo and is deprecated; it will be removed in a future version. Use 'fitness' instead.", DeprecationWarning, stacklevel=2, ) return self.fitness def reset(self) -> "GeneticAlgorithm": """ Reset the solver and return the object. Only use it if you want to have a completely clean sheet. Also, the function is called for every object at instantiation. Note ---- This method resets the internal state of the genetic algorithm, including the population and champion. Use this method when you want to start a new optimization process from scratch. """ self._evolver = self.evolver() self._evolver.send(None) self._champion = None self._state = OptimizerState() self._status = GeneticAlgorithm.Status.INITIALIZED return self def set_solution_params(self, **kwargs) -> "GeneticAlgorithm": """ Set the hyperparameters of the algorithm. Parameters ---------- p_c: float, Optional Probability of crossover. p_m: float, Optional Probability of mutation. maxiter: int, Optional Maximum number of iterations. miniter: int, Optional Minimum number of iterations. elitism: float or int, Optional Determines the portion of the population designated as elite, which automatically survives to the next generation. If less than or equal to 1, it specifies a fraction of the population. If greater than 1, it indicates the exact number of individuals to be selected as elite. A value of 1 assures that the reigning champion is always preserved. To turn this off, set the value to None. maxage: int, Optional Maximum age of the champion. minimize: bool, Optional If True, the objective function is minimized. Default is False. """ if "p_c" in kwargs: self.p_c = kwargs["p_c"] if "p_m" in kwargs: self.p_m = kwargs["p_m"] if "maxiter" in kwargs: self.maxiter = kwargs["maxiter"] if "miniter" in kwargs: self.miniter = kwargs["miniter"] if "elitism" in kwargs: self.elitism = kwargs["elitism"] if "maxage" in kwargs: self.maxage = kwargs["maxage"] if "minimize" in kwargs: self._minimize = kwargs["minimize"] if isinstance(self.elitism, (int, float)): if self.elitism <= 0: raise ValueError("'elitism' must be greater than 0") if self.elitism > 1: if not isinstance(self.elitism, int): raise ValueError("'elitism' must be an integer if greater than 1") if self.elitism >= self.nPop: raise ValueError("'elitism' must be less than 'nPop'") if self.miniter > self.maxiter: raise ValueError("'maxiter' must be greater than 'miniter'") self._celebrate_op = operator.lt if self._minimize else operator.gt return self def evolver(self) -> Iterable: """Return a generator that can be used to manually control evolutions.""" self.genotypes = self.populate() _ = yield yield self.genotypes while True: genotypes = self.select() self.genotypes = self.populate(genotypes) yield self.genotypes def evolve(self, cycles: int = 1) -> Iterable: """Perform a certain number of cycles of evolution and return the genotypes.""" for _ in range(cycles): next(self._evolver) candidate: Genom = self.best_candidate() self._celebrate(candidate) self._state.diversity = self.diversity return self.genotypes def solve(self, recycle: bool = False, **kwargs) -> Genom: """ Solves the problem and returns the champion. .. note:: This class is designed for maximizing the objective function. To minimize it, either negate the objective function or pass ``minimize=True`` when instantiating the class. Parameters ---------- recycle: bool, Optional If True, the leftover resources of the previous calculation are the starting point of the new solution. kwargs: dict Additional parameters to be passed to the :func:`set_solution_params`. Returns ------- :class:`~sigmaepsilon.math.optimize.ga.Genom` The best candidate. """ if not recycle: self.reset() self.set_solution_params(**kwargs) finished = False self._status = GeneticAlgorithm.Status.INITIALIZED try: while not finished: self.evolve(1) self.state.n_iter += 1 min_iter_reached = self.state.n_iter >= self.miniter max_iter_reached = self.state.n_iter >= self.maxiter finished = ( self.stopping_criteria() or max_iter_reached ) and min_iter_reached if finished: if max_iter_reached: self._status = GeneticAlgorithm.Status.MAX_ITERATIONS_REACHED elif self.stopping_criteria(): self._status = GeneticAlgorithm.Status.CONVERGED self._state.success = True else: # pragma: no cover self._state.success = False except Exception as e: self._state.success = False self._state.message = str(e) self._status = GeneticAlgorithm.Status.ERROR raise e finally: self._state.stage = self._status.value return self.champion def evaluate(self, phenotypes: Iterable | None = None) -> ndarray: """ Evaluate the objective for a list of phenotypes. If the phenotypes are not explicitly specified, the population at hand is evaluated. Parameters ---------- phenotypes: Iterable, Optional The phenotypes the objective function is to be evaluated for. Default is None. Note ---- By default, the objective function is called once per individual in a plain Python loop. Set ``vectorized=True`` at construction if the objective function can consume the whole ``(nPop, dim)`` array of phenotypes at once and return one fitness value per row. Alternatively, set ``n_jobs`` to a value other than 1 to evaluate individuals in parallel worker processes (requires a picklable objective function). """ try: phenotypes = self.phenotypes if phenotypes is None else phenotypes if self._is_symbolic_Function: result = self.fnc(phenotypes.T) elif self.vectorized: result = np.asarray(self.fnc(phenotypes), dtype=float) elif self.n_jobs != 1: max_workers = None if self.n_jobs < 0 else self.n_jobs with ProcessPoolExecutor(max_workers=max_workers) as executor: result = np.array( list(executor.map(self.fnc, phenotypes)), dtype=float ) else: result = np.array([self.fnc(x) for x in phenotypes], dtype=float) self.state.n_fev += len(phenotypes) except Exception as e: # pragma: no cover result = None raise RuntimeError( "Error during evaluation of the objective function." ) from e return result def best_phenotype(self) -> ndarray: """ Return the best phenotype from the active population. .. note:: The value returned by this method is the phenotype of the best candidate from the active population, but this is not necessarily the best known solution to the optimization problem at hand. If you want to get the reignng champion, use the :func:`champion` property. """ return self.best_candidate().phenotype def best_candidate(self) -> Genom: """Return the Genom of the best candidate in the active population. .. note:: The value returned by this method is the Genom of the best candidate from the active population, but this is not necessarily the best known solution to the optimization problem at hand. If you want to get the reignng champion, use the :func:`champion` property. """ fitness = self.fitness argfunc = np.argmin if self._minimize else np.argmax index = argfunc(fitness) return Genom( phenotype=self.phenotypes[index], genotype=self.genotypes[index], fitness=fitness[index], index=index, ) def _celebrate(self, genom: Genom) -> None: """Celebrate the winner. Currently, this means that the best candidate is added to a history to keep track of the improvements across evolutions. """ if self.champion is None: self._champion = genom self._champion.age = 0 else: has_new_champion = self._celebrate_op(genom, self.champion) if has_new_champion: self._champion = genom self._champion.age = 0 self.state.x = self.champion.phenotype self.state.fun = self.champion.fitness self._champion.age += 1 def divide(self, fitness: ndarray | None = None) -> tuple[ndarray, ndarray]: """Divide population to elit and others. Returns the corresponding index arrays. Parameters ---------- fitness: numpy.ndarray, Optional Fitness values. If not provided, values from the latest evaluation are used. Default is None. Returns ------- list Indices of the members of the elite. list Indices of the members of the others. """ fitness = self.fitness if fitness is None else fitness assert fitness is not None, "No available fitness data detected." if self.elitism is None: return np.array([], dtype=int), np.arange(self.nPop) if self.elitism is not None: argsort = np.argsort(fitness) if not self._minimize: argsort = argsort[::-1] if self.elitism < 1: elit = argsort[: int(self.nPop * self.elitism)] others = argsort[int(self.nPop * self.elitism) :] else: elit = argsort[: self.elitism] others = argsort[self.elitism :] return elit, others def random_parents_generator(self, genotypes: ndarray) -> Generator: """Yield random pairs from a list of genotypes. The implemantation assumes that the length of the input array is a multiple of 2. Parameters ---------- genotypes: numpy.ndarray Genotypes of the parents as a 2d array. Yields ------ numpy.ndarray The first parent. numpy.ndarray The second parent. """ n = len(genotypes) assert n % 2 == 0, "'n' must be a multiple of 2" pool = np.full(n, True) while True: where = np.argwhere(pool == True).flatten() if len(where) < 2: break pair = self.rng.choice(where, 2, replace=False) parent1 = genotypes[pair[0]] parent2 = genotypes[pair[1]] pool[pair] = False yield parent1, parent2 def stopping_criteria(self) -> bool: """Implement a simple stopping criterion. Return True if the current champion is thought to be the best solution and no further progress can be made, or only at a slow rate. The default implementation considers a champion as the winner, if it is the champion for for at least 5 times in a row. This can be dontrolled with the `maxage` parameter when instantiating an instance. """ return self.champion.age > self.maxage def encode(self, phenotypes: ndarray | None = None) -> ndarray: """Turn phenotypes into genotypes. The default implementation is the identity mapping (genotype == phenotype), suitable for representations that don't need a separate encoding, e.g. real-valued genotypes. Override for representations that do, e.g. binary encoding. """ return phenotypes def decode(self, genotypes: ndarray) -> ndarray: """Turn genotypes into phenotypes. The default implementation is the identity mapping (phenotype == genotype), suitable for representations that don't need a separate decoding, e.g. real-valued genotypes. Override for representations that do, e.g. binary encoding. """ return genotypes def populate(self, genotypes: ndarray | None = None) -> ndarray: """Produce a pool of genotypes. This is a representation-specific extension point with no meaningful generic default; override it in a subclass. """ raise NotImplementedError( f"{type(self).__name__} does not implement 'populate'." ) def crossover(self, parent1: ndarray, parent2: ndarray) -> Tuple[ndarray]: """Take in two parents, return two offspring. You'd probably want to use it inside the populator. This is a representation-specific extension point with no meaningful generic default; override it in a subclass. """ raise NotImplementedError( f"{type(self).__name__} does not implement 'crossover'." ) def mutate(self, child: ndarray) -> ndarray: """Take a child in, return a mutant. This is a representation-specific extension point with no meaningful generic default; override it in a subclass. """ raise NotImplementedError(f"{type(self).__name__} does not implement 'mutate'.") def select( self, genotypes: ndarray | None = None, phenotypes: ndarray | None = None ) -> ndarray: """Run :attr:`selection_strategy` over the current population's fitness values. Returns the genotypes of the winners. .. note:: Providing either ``genotypes`` or ``phenotypes`` (or both) explicitly is not currently supported and raises a ``NotImplementedError``; only the default case (both None, operating on the current population) is implemented. """ if (genotypes is not None) or (phenotypes is not None): raise NotImplementedError( "Selection with either 'genotypes' or 'phenotypes' (or both) provided is " "not implemented. This branch is reached when at least one of these " "arguments is given to 'select', but only the default case (both None) " "is currently supported." ) fitness = self.fitness genotypes = self.genotypes winner_indices = self.selection_strategy.select(self, fitness) return np.asarray([genotypes[w] for w in winner_indices])