def test_genalg_problems():
# Attempt to solve various problems
# Assert that the optimizer can find the solutions
optimizer = GenAlg(32)
optimizer.optimize(problems.ackley_binary,
logging_func=lambda *args: optimize._print_fitnesses(
*args, frequency=100))
assert optimizer.solution_found
def test_genalg_problems():
# Attempt to solve various problems
# Assert that the optimizer can find the solutions
optimizer = GenAlg(32)
optimizer.optimize(
problems.ackley_binary,
logging_func=
lambda *args: optimize._print_fitnesses(*args, frequency=100))
assert optimizer.solution_found
def test_Optimizer_optimize_cache_encoded_True_cache_solution_False():
"""Should only cache decoded solutions if True."""
# After calling Optimizer._get_fitnesses
# __encoded_cache should not be empty
# __solution_cache should be empty
optimizer = GenAlg(2)
# Get fitnesses
optimizer.optimize(SIMPLE_PROBLEM,
max_iterations=1,
cache_encoded=True,
cache_solution=False,
clear_cache=False)
# Assert caches as expected
assert optimizer._Optimizer__encoded_cache != {}
assert optimizer._Optimizer__solution_cache == {}
def test_Optimizer_optimize_cache_encoded_True_cache_solution_False():
"""Should only cache decoded solutions if True."""
# After calling Optimizer._get_fitnesses
# __encoded_cache should not be empty
# __solution_cache should be empty
optimizer = GenAlg(2)
# Get fitnesses
optimizer.optimize(
SIMPLE_PROBLEM,
max_iterations=1,
cache_encoded=True,
cache_solution=False,
clear_cache=False)
# Assert caches as expected
assert optimizer._Optimizer__encoded_cache != {}
assert optimizer._Optimizer__solution_cache == {}
def test_metaoptimize_genalg():
optimizer = GenAlg(32)
prev_hyperparameters = optimizer._get_hyperparameters()
# Test without metaoptimize, save iterations to solution
optimizer.optimize(problems.sphere_binary)
iterations_to_solution = optimizer.iteration
# Test with metaoptimize, assert that iterations to solution is lower
optimizer.optimize_hyperparameters(
problems.sphere_binary,
smoothing=1,
max_iterations=1,
_meta_optimizer=GenAlg(None, population_size=2))
optimizer.optimize(problems.sphere_binary)
assert optimizer._get_hyperparameters() != prev_hyperparameters
def test_Optimizer_meta_optimize_parameter_locks():
# Run meta optimize with locks
# assert that locked parameters did not change
# Only optimize mutation chance
parameter_locks = [
'_population_size', '_crossover_chance', '_selection_function',
'_crossover_function'
]
my_genalg = GenAlg(2)
original = copy.deepcopy(my_genalg)
# Low smoothing for faster performance
my_genalg.optimize_hyperparameters(
SIMPLE_PROBLEM, parameter_locks=parameter_locks, smoothing=1)
# Check that mutation chance changed
assert my_genalg._mutation_chance != original._mutation_chance
# And all others stayed the same
for parameter in parameter_locks:
assert getattr(my_genalg, parameter) == getattr(original, parameter)
def test_Optimizer_meta_optimize_parameter_locks():
# Run meta optimize with locks
# assert that locked parameters did not change
# Only optimize mutation chance
parameter_locks = [
'_population_size', '_crossover_chance', '_selection_function',
'_crossover_function'
]
my_genalg = GenAlg(2)
original = copy.deepcopy(my_genalg)
# Low smoothing for faster performance
my_genalg.optimize_hyperparameters(SIMPLE_PROBLEM,
parameter_locks=parameter_locks,
smoothing=1)
# Check that mutation chance changed
assert my_genalg._mutation_chance != original._mutation_chance
# And all others stayed the same
for parameter in parameter_locks:
assert getattr(my_genalg, parameter) == getattr(original, parameter)
def test_metaoptimize_gsa():
optimizer = GSA(2, [-5.0] * 2, [5.0] * 2)
prev_hyperparameters = optimizer._get_hyperparameters()
# Test without metaoptimize, save iterations to solution
optimizer.optimize(problems.sphere_real)
iterations_to_solution = optimizer.iteration
# Test with metaoptimize, assert that iterations to solution is lower
optimizer.optimize_hyperparameters(problems.sphere_real,
smoothing=1,
max_iterations=1,
_meta_optimizer=GenAlg(
None, population_size=2))
optimizer.optimize(problems.sphere_real)
assert optimizer._get_hyperparameters() != prev_hyperparameters
def test_metaoptimize_crossentropy():
optimizer = crossentropy.CrossEntropy(32)
prev_hyperparameters = optimizer._get_hyperparameters()
# Test without metaoptimize, save iterations to solution
optimizer.optimize(problems.sphere_binary)
iterations_to_solution = optimizer.iteration
# Test with metaoptimize, assert that iterations to solution is lower
optimizer.optimize_hyperparameters(problems.sphere_binary,
smoothing=1,
max_iterations=1,
_meta_optimizer=GenAlg(
None, population_size=2))
optimizer.optimize(problems.sphere_binary)
assert optimizer._get_hyperparameters() != prev_hyperparameters
def test_genalg_sphere_tournament_with_diversity():
_check_optimizer(
GenAlg(32,
selection_function=functools.partial(
gaoperators.tournament_selection, diversity_weight=1.0)))
def test_genalg_chromosome_size_eq_1():
"""Regression test for chromosome_size == 1 edge case."""
optimizer = GenAlg(1)
optimizer.optimize(VERY_SIMPLE_PROBLEM)
assert optimizer.solution_found
def test_genalg_sphere_defaults():
_check_optimizer(GenAlg(32))
def test_Optimizer_optimize_sphere_max_seconds():
optimizer = GenAlg(32, population_size=10)
optimizer.optimize(problems.sphere_binary,
max_iterations=float('inf'),
max_seconds=10)
assert optimizer.solution_found
def optimize_hyperparameters(self,
problems,
parameter_locks=None,
smoothing=20,
max_iterations=100,
_meta_optimizer=None,
_low_memory=True):
"""Optimize hyperparameters for a given problem.
Args:
parameter_locks: a list of strings, each corresponding to a hyperparamter
that should not be optimized.
problems: Either a single problem, or a list of problem instances,
allowing optimization based on multiple similar problems.
smoothing: int; number of runs to average over for each set of hyperparameters.
max_iterations: The number of iterations to optimize before stopping.
_low_memory: disable performance enhancements to save memory
(they use a lot of memory otherwise).
"""
if smoothing <= 0:
raise ValueError('smoothing must be > 0')
# problems supports either one or many problem instances
if isinstance(problems, collections.Iterable):
for problem in problems:
if not isinstance(problem, Problem):
raise TypeError(
'problem must be Problem instance or list of Problem instances'
)
elif isinstance(problems, Problem):
problems = [problems]
else:
raise TypeError(
'problem must be Problem instance or list of Problem instances'
)
# Copy to avoid permanent modification
meta_parameters = copy.deepcopy(self._hyperparameters)
# First, handle parameter locks, since it will modify our
# meta_parameters dict
locked_values = _parse_parameter_locks(self, meta_parameters,
parameter_locks)
# We need to know the size of our chromosome,
# based on the hyperparameters to optimize
solution_size = _get_hyperparameter_solution_size(meta_parameters)
# We also need to create a decode function to transform the binary solution
# into parameters for the metaheuristic
decode = _make_hyperparameter_decode_func(locked_values,
meta_parameters)
# A master fitness dictionary can be stored for use between calls
# to meta_fitness
if _low_memory:
master_fitness_dict = None
else:
master_fitness_dict = {}
additional_parameters = {
'_optimizer': self,
'_problems': problems,
'_runs': smoothing,
'_master_fitness_dict': master_fitness_dict,
}
META_FITNESS = Problem(
_meta_fitness_func,
decode_function=decode,
fitness_kwargs=additional_parameters)
if _meta_optimizer is None:
# Initialize default meta optimizer
# GenAlg is used because it supports both discrete and continous
# attributes
from optimal import GenAlg
# Create metaheuristic with computed decode function and soltuion
# size
_meta_optimizer = GenAlg(solution_size)
else:
# Adjust supplied metaheuristic for this problem
_meta_optimizer._solution_size = solution_size
# Determine the best hyperparameters with a metaheuristic
best_parameters = _meta_optimizer.optimize(
META_FITNESS, max_iterations=max_iterations)
# Set the hyperparameters inline
self._set_hyperparameters(best_parameters)
# And return
return best_parameters
#-20 * math.exp(-0.2 * math.sqrt(0.5 * (x1**2 + x2**2))) - math.exp(
# 0.5 * (math.cos(2 * math.pi * x1) + math.cos(2 * math.pi * x2))) + 20 + math.e
# You can prematurely stop the metaheuristic by returning True
# as the second return value
# Here, we consider the problem solved if the output is <= 0.01
finished = output <= 0.01
# Because this function is trying to minimize the output,
# a smaller output has a greater fitness
fitness = 1 / output
# First return argument must be a real number
# The higher the number, the better the solution
# Second return argument is a boolean, and optional
return fitness, finished
# Define a problem instance to optimize
# We can optionally include a decode function
# The optimizer will pass the decoded solution into your fitness function
# Additional fitness function and decode function parameters can also be added
ackley = Problem(ackley_fitness, decode_function=decode_ackley)
# Create a genetic algorithm with a chromosome size of 32,
# and use it to solve our problem
my_genalg = GenAlg(32)
best_solution = my_genalg.optimize(ackley)
print best_solution
def test_Optimizer_optimize_solution_correct():
optimizer = GenAlg(2)
assert optimizer.optimize(SIMPLE_PROBLEM) == [1, 1]
def ANFISGA():
anfisga = Problem(anfisga_fitness,
decode_function=decode_anfisga_search_space)
my_genalg = GenAlg(32)
best_solution = my_genalg.optimize(anfisga)
print best_solution
def SVMGA():
svmga = Problem(svmga_fitness, decode_function=decode_svmga_search_space)
my_genalg = GenAlg(32)
best_solution = my_genalg.optimize(svmga, max_iterations=1000000)
print best_solution
def test_Optimizer_optimize_solution_correct():
optimizer = GenAlg(2)
assert optimizer.optimize(SIMPLE_PROBLEM) == [1, 1]
def test_Optimizer_optimize_sphere_max_seconds():
optimizer = GenAlg(32, population_size=10)
optimizer.optimize(
problems.sphere_binary, max_iterations=float('inf'),
max_seconds=10)
assert optimizer.solution_found
PROBLEMS = [
problems.ackley_binary, problems.levis_binary, problems.eggholder_binary,
problems.table_binary, problems.shaffer_binary, problems.cross_binary
]
def benchmark_multi(optimizer):
"""Benchmark an optimizer configuration on multiple functions."""
# Get our benchmark stats
all_stats = benchmark.compare(optimizer, PROBLEMS, runs=100)
return benchmark.aggregate(all_stats)
# Create the genetic algorithm configurations to compare
# In reality, we would also want to optimize other hyperparameters
ga_onepoint = GenAlg(32, crossover_function=gaoperators.one_point_crossover)
ga_uniform = GenAlg(32, crossover_function=gaoperators.uniform_crossover)
# Run a benchmark for multiple problems, for robust testing
onepoint_stats = benchmark_multi(ga_onepoint)
uniform_stats = benchmark_multi(ga_uniform)
print
print 'One Point'
pprint.pprint(onepoint_stats)
print
print 'Uniform'
pprint.pprint(uniform_stats)
# We can obtain an easier comparison by performing another aggregate step
aggregate_stats = benchmark.aggregate({
def optimize_hyperparameters(self,
problems,
parameter_locks=None,
smoothing=20,
max_iterations=100,
_meta_optimizer=None,
_low_memory=True):
"""Optimize hyperparameters for a given problem.
Args:
parameter_locks: a list of strings, each corresponding to a hyperparamter
that should not be optimized.
problems: Either a single problem, or a list of problem instances,
allowing optimization based on multiple similar problems.
smoothing: int; number of runs to average over for each set of hyperparameters.
max_iterations: The number of iterations to optimize before stopping.
_low_memory: disable performance enhancements to save memory
(they use a lot of memory otherwise).
"""
if smoothing <= 0:
raise ValueError('smoothing must be > 0')
# problems supports either one or many problem instances
if isinstance(problems, collections.Iterable):
for problem in problems:
if not isinstance(problem, Problem):
raise TypeError(
'problem must be Problem instance or list of Problem instances'
)
elif isinstance(problems, Problem):
problems = [problems]
else:
raise TypeError(
'problem must be Problem instance or list of Problem instances'
)
# Copy to avoid permanent modification
meta_parameters = copy.deepcopy(self._hyperparameters)
# First, handle parameter locks, since it will modify our
# meta_parameters dict
locked_values = _parse_parameter_locks(self, meta_parameters,
parameter_locks)
# We need to know the size of our chromosome,
# based on the hyperparameters to optimize
solution_size = _get_hyperparameter_solution_size(meta_parameters)
# We also need to create a decode function to transform the binary solution
# into parameters for the metaheuristic
decode = _make_hyperparameter_decode_func(locked_values,
meta_parameters)
# A master fitness dictionary can be stored for use between calls
# to meta_fitness
if _low_memory:
master_fitness_dict = None
else:
master_fitness_dict = {}
additional_parameters = {
'_optimizer': self,
'_problems': problems,
'_runs': smoothing,
'_master_fitness_dict': master_fitness_dict,
}
META_FITNESS = Problem(
_meta_fitness_func,
decode_function=decode,
fitness_kwargs=additional_parameters)
if _meta_optimizer is None:
# Initialize default meta optimizer
# GenAlg is used because it supports both discrete and continous
# attributes
from optimal import GenAlg
# Create metaheuristic with computed decode function and soltuion
# size
_meta_optimizer = GenAlg(solution_size)
else:
# Adjust supplied metaheuristic for this problem
_meta_optimizer._solution_size = solution_size
# Determine the best hyperparameters with a metaheuristic
best_parameters = _meta_optimizer.optimize(
META_FITNESS, max_iterations=max_iterations)
# Set the hyperparameters inline
self._set_hyperparameters(best_parameters)
# And return
return best_parameters
print('lr : ' + str(learning_rate) + ', hidden neuron : ' + str(neuron))
output = rmse
finished = output <= err_threshold
#finished = output <= 0.01
fitness = 1 / output
print(finished)
print(fitness)
return fitness, finished
# In[10]:
#Pipeine 3 - FS->ANN->GA
ann_ml = Problem(ann_fs_fitness, decode_function=decode_param)
my_genalg = GenAlg(32, mutation_chance=0.1, crossover_chance=0.8)
start_time = time.time()
best_solution = my_genalg.optimize(ann_ml, max_iterations=15, n_processes=1)
elapsed_time = time.time() - start_time
print(time.strftime("%H:%M:%S", time.gmtime(elapsed_time)))
print best_solution
# In[15]:
print best_solution
# In[11]:
#Pipeine 4 - ANN->GA
ann_ml = Problem(ann_fitness, decode_function=decode_param)
def test_genalg_sphere_stochastic_selection():
# Needs higher population size to consistently succeed
_check_optimizer(
GenAlg(32,
population_size=40,
selection_function=gaoperators.stochastic_selection))
def test_Optimizer_optimize_parallel():
optimzier = GenAlg(2)
optimzier.optimize(SIMPLE_PROBLEM, n_processes=random.randint(2, 4))
assert optimzier.solution_found
def test_Optimizer_optimize_parallel():
optimzier = GenAlg(2)
optimzier.optimize(SIMPLE_PROBLEM, n_processes=random.randint(2, 4))
assert optimzier.solution_found
def test_genalg_chromosome_size_eq_1():
"""Regression test for chromosome_size == 1 edge case."""
optimizer = GenAlg(1)
optimizer.optimize(VERY_SIMPLE_PROBLEM)
assert optimizer.solution_found