def __init__(self, code, *args):
self.code = code
from pyec.util.registry import BENCHMARKS
alpha, bm1, bm2 = code.split("_")
self.alpha = float(alpha)
self.bm1 = BENCHMARKS.load(bm1)
self.bm2 = BENCHMARKS.load(bm2)
def __init__(self, code, *args):
self.code = code
from pyec.util.registry import BENCHMARKS
alpha, bm1, bm2 = code.split("_")
self.alpha = float(alpha)
self.bm1 = BENCHMARKS.load(bm1)
self.bm2 = BENCHMARKS.load(bm2)
def run(self, fitness=None, history=None, extraArgs=None, **kwargs):
"""
Run this :class:`PopulationDistribution` instance to maximize a fitness function.
After running this method, the property ``PopulationDistribution.trainer`` will contain the :class:`Trainer` object used to optimize the function.
:param fitness: The fitness function (objective) to be maximized.
If ``None``, then the function will be looked up from
``pyec.util.registry.BENCHMARKS`` based on the
``function`` property of the :class:`Config`.
:type fitness: any callable object with a single argument, or ``None``
:param history: A history object to extend, or ``None`` to create
a new history from the class in the ``history``
property of the :class:`Config` object.
:type history: :class:`History` or ``None``
:param extraArgs: Any extra args to be passed to
``pyec.util.registry.BENCHMARKS``.
:type extraArgs: list
:returns: The :class:`History for the object
"""
# Set up the fitness if necessary
if fitness is None:
if extraArgs is None:
extraArgs = []
fitness = BENCHMARKS.load(self.config.function, *extraArgs)
try:
fitness.algorithm = self
fitness.config = self.config
except Exception:
pass
self.config.fitness = fitness
# Set up history
if history is None:
history = self.config.history(self.config)
if not self.compatible(history):
name = history.__class__.__name__
raise ValueError("Incompatible history class {0}".format(name))
self.update(history, fitness)
# get the sample
pop = self.population(self.config.populationSize)
# update the history
self.history.update(pop, fitness, self.config.space)
return self.history
def run(self, fitness=None, history=None, extraArgs=None, **kwargs):
"""
Run this :class:`PopulationDistribution` instance to maximize a fitness function.
After running this method, the property ``PopulationDistribution.trainer`` will contain the :class:`Trainer` object used to optimize the function.
:param fitness: The fitness function (objective) to be maximized.
If ``None``, then the function will be looked up from
``pyec.util.registry.BENCHMARKS`` based on the
``function`` property of the :class:`Config`.
:type fitness: any callable object with a single argument, or ``None``
:param history: A history object to extend, or ``None`` to create
a new history from the class in the ``history``
property of the :class:`Config` object.
:type history: :class:`History` or ``None``
:param extraArgs: Any extra args to be passed to
``pyec.util.registry.BENCHMARKS``.
:type extraArgs: list
:returns: The :class:`History for the object
"""
# Set up the fitness if necessary
if fitness is None:
if extraArgs is None:
extraArgs = []
fitness = BENCHMARKS.load(self.config.function, *extraArgs)
try:
fitness.algorithm = self
fitness.config = self.config
except Exception:
pass
self.config.fitness = fitness
# Set up history
if history is None:
history = self.config.history(self.config)
if not self.compatible(history):
name = history.__class__.__name__
raise ValueError("Incompatible history class {0}".format(name))
self.update(history, fitness)
# get the sample
pop = self.population(self.config.populationSize)
# update the history
self.history.update(pop, fitness, self.config.space)
return self.history
def __getitem__(self, pair):
"""Call ``update`` by splitting out the pair into a :class:`History` and
a fitness function. Returns the optimizer for use as a continuation.
:param pair: A tuple with the history and fitness.
:type pair: A ``tuple`` of a :class:`History` and a callable object
:returns: This optimizer (``self``)
"""
history, fitness = pair
# Set up history
if history is None:
history = self.config.history(self.config)
if not self.compatible(history):
err = ("Got an incompatible history in __getitem__; "
"expected [history,fitness]")
raise ValueError(err)
# Set up the fitness if necessary
if fitness is None:
fitness = self.config.function
if isinstance(fitness, basestring):
fitness = BENCHMARKS.load(fitness)
try:
fitness.algorithm = self
fitness.config = self.config
except Exception:
pass
self.config.fitness = fitness
if not inspect.isfunction(fitness) and not hasattr(
fitness, '__call__'):
err = ("Second object in __getitem__ is not a function; "
"expected [history,fitness]")
raise ValueError(err)
self.update(history, fitness)
return self
def __getitem__(self, pair):
"""Call ``update`` by splitting out the pair into a :class:`History` and
a fitness function. Returns the optimizer for use as a continuation.
:param pair: A tuple with the history and fitness.
:type pair: A ``tuple`` of a :class:`History` and a callable object
:returns: This optimizer (``self``)
"""
history, fitness = pair
# Set up history
if history is None:
history = self.config.history(self.config)
if not self.compatible(history):
err = ("Got an incompatible history in __getitem__; "
"expected [history,fitness]")
raise ValueError(err)
# Set up the fitness if necessary
if fitness is None:
fitness = self.config.function
if isinstance(fitness, basestring):
fitness = BENCHMARKS.load(fitness)
try:
fitness.algorithm = self
fitness.config = self.config
except Exception:
pass
self.config.fitness = fitness
if not inspect.isfunction(fitness) and not hasattr(fitness, '__call__'):
err = ("Second object in __getitem__ is not a function; "
"expected [history,fitness]")
raise ValueError(err)
self.update(history, fitness)
return self
def optimize(optimizer, func, dimension=5, population=25, generations=100,**kwargs):
"""
Configure and run an optimizer on a function.
By default the function will be minimize, but maximization can be performed by setting the keyword argument *minimize* to ``False``.
Benchmark functions can be optimized by name. The following names are supported:
- ackley -- A checker-board like oscillator, minimum is -13.37 in 5 dimensions.
- ackley2 -- Exponentiated and centered version of ackley, minimum is 0 at 0.
- griewank -- Oscillator with large scale, minimum at 0.
- langerman -- Sparse, rough, multi-modal. Minimum is 0.98 in five dimensions.
- rosenbrock -- Standard benchmark.
- rastrigin -- Oscillator. Minimum at
- salomon -- Ring oscillation. Minimum 0 at 0.
- schwefel -- Deceptive multimodal function. Minimum is -418 on (-512,512).
- shekel2 -- Shekel's foxholes, modified. Minimum is -10.4 in five dimensions.
- sphere -- A spherical paraboloid, minimum is 0 at 0
- whitley -- Complex, fractal like shape with small relevant area. Minimum is 0.0.
- weierstrass -- Everywhere continuous, nowhere differentiable; minimum is 0 at 0.
:param optimizer: A :class:`PopulationDistribution` subclass
:type optimizer: ``class``
:param func: The function to be optimized, or a lookup key for a benchmark.
:type func: any callable object or str
:param dimension: The vector dimension in the search domain
:type dimension: int
:param population: The population size (sample size) for the optimizer.
:type population: int
:param generations: The number of populations to build (number of samples) during optimization.
:type generations: int
:returns: A tuple (best solution, best value) where the first element is the *best solution* located during optimization and *best value* is the value of the function at *best solution*.
Keyword arguments:
* minimize -- Whether to minimize the function, otherwise maximize; default is True.
* initial -- A callable (no arguments) that returns random starting points for the initial distribution of the optimizer.
* display -- Show progress information once every second.
* constraint -- A :class:`Boundary` object implementing a constraint region (default is unconstrained).
"""
space = ("constraint" in kwargs and kwargs["constraint"]
or Euclidean(dim=dimension))
config = {
"minimize":True,
"space":space,
"populationSize":population
}
config.update(kwargs)
if isinstance(func, basestring):
from pyec.util.registry import BENCHMARKS
func = BENCHMARKS.load(func)
#if config["minimize"]:
# h = func
# func = lambda x: -h(x)
if config["minimize"]:
optfunc = lambda x: -func(x)
else:
optfunc = func
config = Config(**config)
alg = (optimizer[config] << generations)()
pop = alg[None, optfunc]()
alg.history.update(pop, optfunc, space, alg)
return alg.history.best()
def optimize(optimizer,
func,
dimension=5,
population=25,
generations=100,
**kwargs):
"""
Configure and run an optimizer on a function.
By default the function will be minimize, but maximization can be performed by setting the keyword argument *minimize* to ``False``.
Benchmark functions can be optimized by name. The following names are supported:
- ackley -- A checker-board like oscillator, minimum is -13.37 in 5 dimensions.
- ackley2 -- Exponentiated and centered version of ackley, minimum is 0 at 0.
- griewank -- Oscillator with large scale, minimum at 0.
- langerman -- Sparse, rough, multi-modal. Minimum is 0.98 in five dimensions.
- rosenbrock -- Standard benchmark.
- rastrigin -- Oscillator. Minimum at
- salomon -- Ring oscillation. Minimum 0 at 0.
- schwefel -- Deceptive multimodal function. Minimum is -418 on (-512,512).
- shekel2 -- Shekel's foxholes, modified. Minimum is -10.4 in five dimensions.
- sphere -- A spherical paraboloid, minimum is 0 at 0
- whitley -- Complex, fractal like shape with small relevant area. Minimum is 0.0.
- weierstrass -- Everywhere continuous, nowhere differentiable; minimum is 0 at 0.
:param optimizer: A :class:`PopulationDistribution` subclass
:type optimizer: ``class``
:param func: The function to be optimized, or a lookup key for a benchmark.
:type func: any callable object or str
:param dimension: The vector dimension in the search domain
:type dimension: int
:param population: The population size (sample size) for the optimizer.
:type population: int
:param generations: The number of populations to build (number of samples) during optimization.
:type generations: int
:returns: A tuple (best solution, best value) where the first element is the *best solution* located during optimization and *best value* is the value of the function at *best solution*.
Keyword arguments:
* minimize -- Whether to minimize the function, otherwise maximize; default is True.
* initial -- A callable (no arguments) that returns random starting points for the initial distribution of the optimizer.
* display -- Show progress information once every second.
* constraint -- A :class:`Boundary` object implementing a constraint region (default is unconstrained).
"""
space = ("constraint" in kwargs and kwargs["constraint"]
or Euclidean(dim=dimension))
config = {"minimize": True, "space": space, "populationSize": population}
config.update(kwargs)
if isinstance(func, basestring):
from pyec.util.registry import BENCHMARKS
func = BENCHMARKS.load(func)
#if config["minimize"]:
# h = func
# func = lambda x: -h(x)
if config["minimize"]:
optfunc = lambda x: -func(x)
else:
optfunc = func
config = Config(**config)
alg = (optimizer[config] << generations)()
pop = alg[None, optfunc]()
alg.history.update(pop, optfunc, space, alg)
return alg.history.best()