def run(self, seed):
"""
Trains this classifier.
:param seed: seed used to partition training set at every generation. The (sub)sets will be constant throughout
all the evolutionary process, allowing a direct comparison between individuals from different generations.
:type seed: int
:rtype: tuple
:return: a tuple containing two individuals.Individual objects, where the first individual is the best solution
(according to fitness) found throughout all the evolutionary process, and the second individual the best solution
from the last generation.
"""
# Statistics computation
stats = tools.Statistics(lambda ind: ind.fitness)
for stat_name, stat_func in PBILLogger.population_operators:
stats.register(stat_name, stat_func)
# An object that keeps track of the best individual found so far.
self._hof = HallOfFame(maxsize=1) # type: HallOfFame
best_last, logbook = self.__run__(seed=seed,
ngen=self.n_generations,
stats=stats,
verbose=True)
best_overall = self._hof[0] # type: Individual
self._hof = None
gc.collect()
return best_overall, best_last
def __init__(self,
weights=(),
nr_of_generations=None,
crossover_rate=None,
mutation_rate=None,
pop_size=None,
caching=False):
'''
:param weights: the weights on the outcomes
:param nr_of_generations: the number of generations
:param crossover_rate: the crossover rate
:param mutation_rate: the mutation rate
:param caching: parameter controling wether a list of tried solutions
should be kept
'''
self.stats = []
if len(weights)>1:
self.hall_of_fame = ParetoFront(similar=compare)
else:
self.hall_of_fame = HallOfFame(pop_size)
if caching:
self.tried_solutions = TriedSolutions()
self.change = []
self.weights = weights
self.nr_of_generations = nr_of_generations
self.crossover_rate = crossover_rate
self.mutation_rate=mutation_rate
self.precision = "{0:.%if}" % 2
def __init__(self, **kwargs):
## TODO: add ability to determine if evolution has stopped
## TODO: add archive
## TODO: add generating and collecting different statistics
## TODO: add logging
# create initial populations of all species
# taking part in coevolution
self.kwargs = deepcopy(kwargs)
self.ENV = kwargs["env"]
self.SPECIES = kwargs["species"]
self.INTERACT_INDIVIDUALS_COUNT = kwargs["interact_individuals_count"]
self.GENERATIONS = kwargs["generations"]
self.solstat = kwargs.get("solstat", lambda sols: {})
#choose = kwargs["operators"]["choose"]
self.build_solutions = kwargs["operators"]["build_solutions"]
self.fitness = kwargs["operators"]["fitness"]
self.assign_credits = kwargs["operators"]["assign_credits"]
self.analyzers = kwargs.get("analyzers", [])
self.USE_CREDIT_INHERITANCE = kwargs.get("use_credit_inheritance", False)
self.HALL_OF_FAME_SIZE = kwargs.get("hall_of_fame_size", 0)
self._result = None
self.stat = tools.Statistics(key=lambda x: x.fitness)
self.stat.register("best", rounddec(numpy.max))
self.stat.register("min", rounddec(numpy.min))
self.stat.register("avg", rounddec(numpy.average))
self.stat.register("std", rounddec(numpy.std))
self.logbook = tools.Logbook()
self.kwargs['logbook'] = self.logbook
## TODO: make processing of population consisting of 1 element uniform
## generate initial population
self.pops = {s: self._generate_k(s.initialize(self.kwargs, s.pop_size)) if not s.fixed
else self._generate_k([s.representative_individual])
for s in self.SPECIES}
## make a copy for logging. TODO: remake it with logbook later.
self.initial_pops = {s.name: deepcopy(pop) for s, pop in self.pops.items()}
## checking correctness
for s, pop in self.pops.items():
sm = sum(p.k for p in pop)
assert sm == self.INTERACT_INDIVIDUALS_COUNT, \
"For specie {0} count doesn't equal to {1}. Actual value {2}".format(s, self.INTERACT_INDIVIDUALS_COUNT, sm)
print("Initialization finished")
self.hall = HallOfFame(self.HALL_OF_FAME_SIZE)
self.kwargs['gen'] = 0
pass
def _select_best(self, pop):
"""Select the best individuals of the population.
:param pop: The population
:type pop: Any iterable type
:return: The best individuals found
:rtype: :py:class:`~deap.tools.HallOfFame` of individuals
"""
hof = HallOfFame(self.pop_size)
hof.update(pop)
return hof
def _setup(self,
refiner,
ngen=NGEN,
ngen_comm=NGEN_COMM,
nswarms=NSWARMS,
**kwargs):
if not self._has_been_setup:
logger.info(
"Setting up the DEAP CMA-ES refinement algorithm (ngen=%d)" %
ngen)
refiner.status.message = "Setting up algorithm..."
# Process some general stuff:
bounds = np.array(refiner.ranges) #@UndefinedVariable
create_individual = self._individual_creator(refiner, bounds)
# Setup strategy:
#TODO make the others random
parents = [None] * nswarms
parents[0] = create_individual(refiner.history.initial_solution)
for i in range(1, nswarms):
parents[i] = create_individual([
random.uniform(bounds[j, 0], bounds[j, 1])
for j in range(len(refiner.history.initial_solution))
])
strategy = SwarmStrategy(
parents=parents,
sigma=1.0 / 10.0,
ranges=bounds,
stop=self._stop,
)
# Toolbox setup:
toolbox = base.Toolbox()
toolbox.register("generate", strategy.generate, create_individual)
toolbox.register("update", strategy.update)
# Hall of fame & stats:
logger.info("Creating hall-off-fame and statistics")
halloffame = HallOfFame(
1, similar=lambda a1, a2: np.all(a1 == a2)
) #PyXRDParetoFront(similar=lambda a1, a2: np.all(a1 == a2))
stats = self._create_stats()
# Create algorithm
self.algorithm = SwarmAlgorithm(toolbox,
halloffame,
stats,
ngen=ngen,
ngen_comm=ngen_comm,
refiner=refiner,
stop=self._stop)
self._has_been_setup = True
return self.algorithm
def main():
pop = toolbox.population(n=rd.pop_size)
stats_cost = tools.Statistics(lambda ind: ind.fitness.values[0])
mstats = tools.MultiStatistics(cost=stats_cost)
mstats.register("min", np.min, axis=0)
mstats.register("median", np.median, axis=0)
mstats.register("max", np.max, axis=0)
hof = HallOfFame(1)
pop, logbook = my_ea.ea(pop,
toolbox,
rd.cxpb,
rd.mutpb,
rd.elitism,
rd.gens,
stats=mstats,
halloffame=hof,
verbose=True)
return pop, mstats, hof, logbook
def __init__(self, bset):
self.bset = bset
pop = [Population.INDIVIDUAL_CLASS(self.bset) for _ in range(self.POPULATION_SIZE)]
super(Population, self).__init__(pop)
self.stats = Statistics(lambda ind: ind.fitness.values)
self.stats.register("avg", np.mean)
self.stats.register("std", np.std)
self.stats.register("min", np.min)
self.stats.register("max", np.max)
self.logbook = Logbook()
self.logbook.header = ['gen'] + self.stats.fields
self.hof = HallOfFame(1)
self.generation = 0
# do an initial evaluation
for ind in self:
ind.fitness.values = ind.evaluate()
def __init__(self, eps):
self.eps = eps
HallOfFame.__init__(self, None)
self.init = False
class BaseLoop(Toolbox):
"""
Base loop for BGP.
Examples::
if __name__ == "__main__":
pset = SymbolSet()
stop = lambda ind: ind.fitness.values[0] >= 0.880963
bl = BaseLoop(pset=pset, gen=10, pop=1000, hall=1, batch_size=40, re_hall=3, \n
n_jobs=12, mate_prob=0.9, max_value=5, initial_min=1, initial_max=2, \n
mutate_prob=0.8, tq=True, dim_type="coef", stop_condition=stop,\n
re_Tree=0, store=False, random_state=1, verbose=True,\n
stats={"fitness_dim_max": ["max"], "dim_is_target": ["sum"]},\n
add_coef=True, inter_add=True, inner_add=False, cal_dim=True, vector_add=False,\n
personal_map=False)
bl.run()
"""
def __init__(self, pset, pop=500, gen=20, mutate_prob=0.5, mate_prob=0.8, hall=1, re_hall=1,
re_Tree=None, initial_min=None, initial_max=3, max_value=5,
scoring=(r2_score,), score_pen=(1,), filter_warning=True, cv=1,
add_coef=True, inter_add=True, inner_add=False, vector_add=False, out_add=False, flat_add=False,
cal_dim=False, dim_type=None, fuzzy=False, n_jobs=1, batch_size=40,
random_state=None, stats=None, verbose=True, migrate_prob=0,
tq=True, store=False, personal_map=False, stop_condition=None, details=False, classification=False,
score_object="y", sub_mu_max=1, limit_type="h_bgp", batch_para=False):
"""
Parameters
----------
pset:SymbolSet
the feature x and target y and others should have been added.
pop: int
number of population.
gen: int
number of generation.
mutate_prob:float
probability of mutate.
mate_prob:float
probability of mate(crossover).
initial_max:int
max initial size of expression when first producing.
initial_min : None,int
min initial size of expression when first producing.
max_value:int
max size of expression.
limit_type: "height" or "length",","h_bgp"
limitation type for max_value, but don't affect initial_max, initial_min.
hall:int,>=1
number of HallOfFame (elite) to maintain.
re_hall:None or int>=2
Notes: only valid when hall
number of HallOfFame to add to next generation.
re_Tree: int
number of new features to add to next generation.
0 is false to add.
personal_map:bool or "auto"
"auto" is using 'premap' and with auto refresh the 'premap' with individual.\n
True is just using constant 'premap'.\n
False is just use the prob of terminals.
scoring: list of Callable, default is [sklearn.metrics.r2_score,]
See Also ``sklearn.metrics``
score_pen: tuple of 1, -1 or float but 0.
>0 : max problem, best is positive, worse -np.inf.
<0 : min problem, best is negative, worse np.inf.
Notes:
if multiply score method, the scores must be turn to same dimension in prepossessing
or weight by score_pen. Because the all the selection are stand on the mean(w_i*score_i)
Examples::
scoring = [r2_score,]
score_pen= [1,]
cv:sklearn.model_selection._split._BaseKFold,int
the shuffler must be False,
default=1 means no cv.
filter_warning:bool
filter warning or not.
add_coef:bool
add coef in expression or not.
inter_add：bool
add intercept constant or not.
inner_add:bool
add inner coefficients or not.
out_add:bool
add out coefficients or not.
flat_add:bool
add flat coefficients or not.
n_jobs:int
default 1, advise 6.
batch_size:int
default 40, depend of machine.
random_state:int
None,int.
cal_dim:bool
escape the dim calculation.
dim_type:Dim or None or list of Dim
"coef": af(x)+b. a,b have dimension,f(x)'s dimension is not dnan. \n
"integer": af(x)+b. f(x) is with integer dimension. \n
[Dim1,Dim2]: f(x)'s dimension in list. \n
Dim: f(x) ~= Dim. (see fuzzy) \n
Dim: f(x) == Dim. \n
None: f(x) == pset.y_dim
fuzzy:bool
choose the dim with same base with dim_type, such as m,m^2,m^3.
stats:dict
details of logbook to show. \n
Map:\n
values
= {"max": np.max, "mean": np.mean, "min": np.mean, "std": np.std, "sum": np.sum}
keys
= {\n
"fitness": just see fitness[0], \n
"fitness_dim_max": max problem, see fitness with demand dim,\n
"fitness_dim_min": min problem, see fitness with demand dim,\n
"dim_is_target": demand dim,\n
"coef": dim is True, coef have dim, \n
"integer": dim is integer, \n
...
}
if stats is None, default is:
for cal_dim=True:
stats = {"fitness_dim_max": ("max",), "dim_is_target": ("sum",)}
for cal_dim=False
stats = {"fitness": ("max",)}
if self-definition, the key is func to get attribute of each ind.
Examples::
def func(ind):
return ind.fitness[0]
stats = {func: ("mean",), "dim_is_target": ("sum",)}
verbose:bool
print verbose logbook or not.
tq:bool
print progress bar or not.
store:bool or path
bool or path.
stop_condition:callable
stop condition on the best ind of hall, which return bool,the true means stop loop.
Examples::
def func(ind):
c = ind.fitness.values[0]>=0.90
return c
details:bool
return expr and predict_y or not.
classification: bool
classification or not.
score_object:
score by y or delta y (for implicit function).
"""
super(BaseLoop, self).__init__()
assert initial_max <= max_value, "the initial size of expression should less than max_value limitation"
if cal_dim:
assert all(
[isinstance(i, Dim) for i in pset.dim_ter_con.values()]), \
"all import dim of pset should be Dim object."
self.details = details
self.max_value = max_value
self.pop = pop
self.gen = gen
self.mutate_prob = mutate_prob
self.mate_prob = mate_prob
self.migrate_prob = migrate_prob
self.verbose = verbose
self.cal_dim = cal_dim
self.re_hall = re_hall
self.re_Tree = re_Tree
self.store = store
self.limit_type = limit_type
self.data_all = []
self.personal_map = personal_map
self.stop_condition = stop_condition
self.population = []
self.rand_state = None
self.random_state = random_state
self.sub_mu_max = sub_mu_max
self.population_next = []
self.cpset = CalculatePrecisionSet(pset, scoring=scoring, score_pen=score_pen,
filter_warning=filter_warning, cal_dim=cal_dim,
add_coef=add_coef, inter_add=inter_add, inner_add=inner_add,
vector_add=vector_add, out_add=out_add, flat_add=flat_add, cv=cv,
n_jobs=n_jobs, batch_size=batch_size, tq=tq,
fuzzy=fuzzy, dim_type=dim_type, details=details,
classification=classification, score_object=score_object,
batch_para=batch_para
)
Fitness_ = newclass.create("Fitness_", Fitness, weights=score_pen)
self.PTree = newclass.create("PTrees", SymbolTree, fitness=Fitness_)
# def produce
if initial_min is None:
initial_min = 2
self.register("genGrow", genGrow, pset=self.cpset, min_=initial_min, max_=initial_max + 1,
personal_map=self.personal_map)
self.register("genFull", genFull, pset=self.cpset, min_=initial_min, max_=initial_max + 1,
personal_map=self.personal_map)
self.register("genHalf", genGrow, pset=self.cpset, min_=initial_min, max_=initial_max + 1,
personal_map=self.personal_map)
self.register("gen_mu", genGrow, min_=1, max_=self.sub_mu_max + 1, personal_map=self.personal_map)
# def selection
self.register("select", selTournament, tournsize=2)
self.register("selKbestDim", selKbestDim,
dim_type=self.cpset.dim_type, fuzzy=self.cpset.fuzzy)
self.register("selBest", selBest)
self.register("mate", cxOnePoint)
# def mutate
self.register("mutate", mutUniform, expr=self.gen_mu, pset=self.cpset)
self.decorate("mate", staticLimit(key=operator.attrgetter(limit_type), max_value=self.max_value))
self.decorate("mutate", staticLimit(key=operator.attrgetter(limit_type), max_value=self.max_value))
if stats is None:
if cal_dim:
if score_pen[0] > 0:
stats = {"fitness_dim_max": ("max",), "dim_is_target": ("sum",)}
else:
stats = {"fitness_dim_min": ("min",), "dim_is_target": ("sum",)}
else:
if score_pen[0] > 0:
stats = {"fitness": ("max",)}
else:
stats = {"fitness": ("min",)}
self.stats = Statis_func(stats=stats)
logbook = Logbook()
logbook.header = ['gen'] + (self.stats.fields if self.stats else [])
self.logbook = logbook
if hall is None:
hall = 1
self.hall = HallOfFame(hall)
if re_hall is None:
self.re_hall = None
else:
if re_hall == 1 or re_hall == 0:
print("re_hall should more than 1")
re_hall = 2
assert re_hall >= hall, "re_hall should more than hall"
self.re_hall = HallOfFame(re_hall)
def varAnd(self, *arg, **kwargs):
return varAnd(*arg, **kwargs)
def to_csv(self, data_all):
"""store to csv"""
if self.store:
if isinstance(self.store, str):
path = self.store
else:
path = os.getcwd()
file_new_name = "_".join((str(self.pop), str(self.gen),
str(self.mutate_prob), str(self.mate_prob),
str(time.time())))
try:
st = Store(path)
st.to_csv(data_all, file_new_name, transposition=True)
print("store data to ", path, file_new_name)
except (IOError, PermissionError):
st = Store(os.getcwd())
st.to_csv(data_all, file_new_name, transposition=True)
print("store data to ", os.getcwd(), file_new_name)
def maintain_halls(self, population):
"""maintain the best p expression"""
if self.re_hall is not None:
maxsize = max(self.hall.maxsize, self.re_hall.maxsize)
if self.cal_dim:
inds_dim = self.selKbestDim(population, maxsize)
else:
inds_dim = self.selBest(population, maxsize)
self.hall.update(inds_dim)
self.re_hall.update(inds_dim)
sole_inds = [i for i in self.re_hall.items if i not in inds_dim]
inds_dim.extend(sole_inds)
else:
if self.cal_dim:
inds_dim = self.selKbestDim(population, self.hall.maxsize)
else:
inds_dim = self.selBest(population, self.hall.maxsize)
self.hall.update(inds_dim)
inds_dim = []
inds_dim = copy.deepcopy(inds_dim)
return inds_dim
def re_add(self):
"""add the expression as a feature"""
if self.hall.items and self.re_Tree:
it = self.hall.items
indo = it[random.choice(len(it))]
ind = copy.deepcopy(indo)
inds = ind.depart()
if not inds:
pass
else:
inds = [self.cpset.calculate_detail(indi) for indi in inds]
le = min(self.re_Tree, len(inds))
indi = inds[random.choice(le)]
self.cpset.add_tree_to_features(indi)
def re_fresh_by_name(self, *arr):
re_name = ["mutate", "genGrow", "genFull", "genHalf"]
if len(arr) > 0:
re_name.extend(arr)
self.refresh(re_name, pset=self.cpset)
# for i in re_name + ["mate"]: # don‘t del this
# self.decorate(i, staticLimit(key=operator.attrgetter("height"), max_value=2 * (self.max_value + 1)))
def top_n(self, n=10, gen=-1, key="value", filter_dim=True, ascending=False):
"""
Return the best n results.
Note:
Only valid in ``store=True``.
Parameters
----------
n:int
n.
gen:
the generation, default is -1.
key: str
sort keys, default is "values".
filter_dim:
filter no-dim expressions or not.
ascending:
reverse.
Returns
-------
top n results.
pd.DataFrame
"""
import pandas as pd
if self.store == "False":
raise TypeError("Only valid with store=True")
data = self.data_all
data = pd.DataFrame(data)
if gen == -1:
gen = max(data["gen"])
data = data[data["gen"] == gen]
if filter_dim:
data = data[data["dim_score"] == 1]
data = data.drop_duplicates(['expr'], keep="first")
if key is not None:
data[key] = data[key].str.replace("(", "")
data[key] = data[key].str.replace(")", "")
data[key] = data[key].str.replace(",", "")
try:
data[key] = data[key].astype(float)
except ValueError:
raise TypeError("check this key column can be translated into float")
data = data.sort_values(by='value', ascending=ascending).iloc[:n, :]
return data
def check_height_length(self, pop, site=""):
old = len(pop)
if self.limit_type == 'height':
pop = [i for i in pop if i.height <= self.max_value]
elif self.limit_type == 'length':
pop = [i for i in pop if i.length <= self.max_value]
else:
pop = [i for i in pop if i.h_bgp <= self.max_value]
new = len(pop)
if old == new:
pass
else:
if site != "":
print(site)
# raise TypeError
index = random.randint(0, new, old - new)
pop.extend([pop[i] for i in index])
return pop
def run(self, warm_start=False, new_gen=None):
"""
Parameters
----------
warm_start:bool
warm_start from last result.
new_gen:
new generations for warm_startm, default is the initial generations.
"""
# 1.generate###################################################################
if warm_start is False:
random.seed(self.random_state)
population = [self.PTree(self.genHalf()) for _ in range(self.pop)]
gen_i = 0
gen = self.gen
else:
assert self.population_next != []
random.set_state(self.rand_state)
population = self.population_next
gen_i = self.gen_i
self.re_fresh_by_name()
if new_gen:
gen = gen_i + new_gen
else:
gen = gen_i + self.gen
for gen_i in range(gen_i + 1, gen + 1):
population_old = copy.deepcopy(population)
# 2.evaluate###############################################################
invalid_ind_score = self.cpset.parallelize_score(population_old)
for ind, score in zip(population_old, invalid_ind_score):
ind.fitness.values = tuple(score[0])
ind.y_dim = score[1]
ind.dim_score = score[2]
ind.coef_expr = score[3]
ind.coef_pre_y = score[4]
population = population_old
# 3.log###################################################################
# 3.1.log-print##############################
record = self.stats.compile(population) if self.stats else {}
self.logbook.record(gen=gen_i, **record)
if self.verbose:
print(self.logbook.stream)
# 3.2.log-store##############################
if self.store:
datas = [{"gen": gen_i, "name": str(pop_i), "expr": str([pop_i.coef_expr]),
"value": str(pop_i.fitness.values),
"dimension": str(pop_i.y_dim),
"dim_score": pop_i.dim_score} for pop_i in population]
self.data_all.extend(datas)
self.population = copy.deepcopy(population)
# 3.3.log-hall###############################
inds_dim = self.maintain_halls(population)
# 4.refresh################################################################
# 4.1.re_update the premap ##################
if self.personal_map == "auto":
[self.cpset.premap.update(indi, self.cpset) for indi in inds_dim]
# 4.2.re_add_tree and refresh pset###########
if self.re_Tree:
self.re_add()
self.re_fresh_by_name()
# 6.next generation ！！！！#######################################################
# selection and mutate,mate,migration
population = self.select(population, int((1 - self.migrate_prob) * len(population)) - len(inds_dim))
offspring = self.varAnd(population, self, self.mate_prob, self.mutate_prob)
offspring.extend(inds_dim)
migrate_pop = [self.PTree(self.genFull()) for _ in range(int(self.migrate_prob * len(population)))]
population[:] = offspring + migrate_pop
population = self.check_height_length(population)
# 5.break#######################################################
if self.stop_condition is not None:
if self.stop_condition(self.hall.items[0]):
break
# 7 freeze ###################################################
self.rand_state = random.get_state()
self.population_next = population
self.gen_i = gen_i
# final.store#####################################################################
if self.store:
self.to_csv(self.data_all)
self.hall.items = [self.cpset.calculate_detail(indi) for indi in self.hall.items]
return self.hall
def mainPart(x_, y_, pset, pop_n=100, random_seed=1, cxpb=0.8, mutpb=0.1, ngen=5, alpha=1,
tournsize=3, max_value=10, double=False, score=None, **kargs):
"""
Parameters
----------
score
double
x_
y_
pset
pop_n
random_seed
cxpb
mutpb
ngen
alpha
tournsize
max_value
kargs
Returns
-------
"""
max_ = pset.max_
if score is None:
score = [r2_score, explained_variance_score]
random.seed(random_seed)
toolbox = Toolbox()
if isinstance(pset, PrimitiveSet):
PTrees = ExpressionTree
Generate = genHalfAndHalf
mutate = mutNodeReplacement
mate = cxOnePoint
elif isinstance(pset, FixedPrimitiveSet):
PTrees = FixedExpressionTree
Generate = generate_
mate = partial(cxOnePoint_index, pset=pset)
mutate = mutUniForm_index
else:
raise NotImplementedError("get wrong pset")
if double:
creator.create("Fitness_", Fitness, weights=(1.0, 1.0))
else:
creator.create("Fitness_", Fitness, weights=(1.0,))
creator.create("PTrees_", PTrees, fitness=creator.Fitness_)
toolbox.register("generate_", Generate, pset=pset, min_=None, max_=max_)
toolbox.register("individual", initIterate, container=creator.PTrees_, generator=toolbox.generate_)
toolbox.register('population', initRepeat, container=list, func=toolbox.individual)
# def selection
toolbox.register("select_gs", selTournament, tournsize=tournsize)
# def mate
toolbox.register("mate", mate)
# def mutate
toolbox.register("mutate", mutate, pset=pset)
if isinstance(pset, PrimitiveSet):
toolbox.decorate("mate", staticLimit(key=operator.attrgetter("height"), max_value=max_value))
toolbox.decorate("mutate", staticLimit(key=operator.attrgetter("height"), max_value=max_value))
# def elaluate
toolbox.register("evaluate", calculate, pset=pset, x=x_, y=y_, score_method=score[0], **kargs)
toolbox.register("evaluate2", calculate, pset=pset, x=x_, y=y_, score_method=score[1], **kargs)
stats1 = Statistics(lambda ind: ind.fitness.values[0])
stats = MultiStatistics(score1=stats1, )
stats.register("avg", np.mean)
stats.register("max", np.max)
pop = toolbox.population(n=pop_n)
haln = 5
hof = HallOfFame(haln)
if double:
population, logbook = multiEaSimple(pop, toolbox, cxpb=cxpb, mutpb=mutpb, ngen=ngen, stats=stats, alpha=alpha,
halloffame=hof, pset=pset)
else:
population, logbook = eaSimple(pop, toolbox, cxpb=cxpb, mutpb=mutpb, ngen=ngen, stats=stats,
halloffame=hof, pset=pset)
return population, logbook, hof
def __init__(self, similar=compare):
self.similar = similar
HallOfFame.__init__(self, None)
class Population(list):
"""
A collection of individuals
"""
INDIVIDUAL_CLASS = Individual
POPULATION_SIZE = 100
CLONE_BEST = 5
MAX_MATE_ATTEMPTS = 10
MATE_MUTATE_CLONE = (80, 18, 2)
def __init__(self, bset):
self.bset = bset
pop = [Population.INDIVIDUAL_CLASS(self.bset) for _ in range(self.POPULATION_SIZE)]
super(Population, self).__init__(pop)
self.stats = Statistics(lambda ind: ind.fitness.values)
self.stats.register("avg", np.mean)
self.stats.register("std", np.std)
self.stats.register("min", np.min)
self.stats.register("max", np.max)
self.logbook = Logbook()
self.logbook.header = ['gen'] + self.stats.fields
self.hof = HallOfFame(1)
self.generation = 0
# do an initial evaluation
for ind in self:
ind.fitness.values = ind.evaluate()
def select(self, k):
"""Probablistic select *k* individuals among the input *individuals*. The
list returned contains references to the input *individuals*.
:param k: The number of individuals to select.
:returns: A list containing k individuals.
The individuals returned are randomly selected from individuals according
to their fitness such that the more fit the individual the more likely
that individual will be chosen. Less fit individuals are less likely, but
still possibly, selected.
"""
# adjusted pop is a list of tuples (adjusted fitness, individual)
adjusted_pop = [(1.0 / (1.0 + i.fitness.values[0]), i) for i in self]
# normalised_pop is a list of tuples (float, individual) where the float indicates
# a 'share' of 1.0 that the individual deserves based on it's fitness relative to
# the other individuals. It is sorted so the best chances are at the front of the list.
denom = sum([fit for fit, ind in adjusted_pop])
normalised_pop = [(fit / denom, ind) for fit, ind in adjusted_pop]
normalised_pop = sorted(normalised_pop, key=lambda i: i[0], reverse=True)
# randomly select with a fitness bias
# FIXME: surely this can be optimized?
selected = []
for x in range(k):
rand = random.random()
accumulator = 0.0
for share, ind in normalised_pop:
accumulator += share
if rand <= accumulator:
selected.append(ind)
break
if len(selected) == 1:
return selected[0]
else:
return selected
def evolve(self):
"""
Evolve this population by one generation
"""
self.logbook.record(gen=self.generation, **self.stats.compile(self))
self.hof.update(self)
print(self.logbook.stream)
# the best x of the population are cloned directly into the next generation
offspring = self[:self.CLONE_BEST]
# rest of the population clone, mate, or mutate at random
for idx in range(len(self) - self.CLONE_BEST):
# decide how to alter this individual
rand = random.randint(0,100)
for _ in range(0, self.MAX_MATE_ATTEMPTS):
try:
if rand < self.MATE_MUTATE_CLONE[0]: # MATE/CROSSOVER
receiver, contributor = self.select(2)
child = receiver.clone()
child.mate(contributor)
break
elif rand < (self.MATE_MUTATE_CLONE[0] + self.MATE_MUTATE_CLONE[1]): # MUTATE
ind = self.select(1)
child = ind.clone()
child.mutate()
break
else:
child = self.select(1).clone()
break
except BirthError:
continue
# generate new blood when reproduction fails so badly
else:
child = Population.INDIVIDUAL_CLASS(self.bset)
offspring.append(child)
self[:] = offspring
self.generation += 1
# evaluate every individual and sort
for ind in self:
if not len(ind.fitness.values):
ind.fitness.values = ind.evaluate()
self.sort(key=lambda i: i.fitness.values[0])
def __init__(self):
HallOfFame.__init__(self, None)
class CoevolutionGA:
def __init__(self, **kwargs):
## TODO: add ability to determine if evolution has stopped
## TODO: add archive
## TODO: add generating and collecting different statistics
## TODO: add logging
# create initial populations of all species
# taking part in coevolution
self.kwargs = deepcopy(kwargs)
self.ENV = kwargs["env"]
self.SPECIES = kwargs["species"]
self.INTERACT_INDIVIDUALS_COUNT = kwargs["interact_individuals_count"]
self.GENERATIONS = kwargs["generations"]
self.solstat = kwargs.get("solstat", lambda sols: {})
#choose = kwargs["operators"]["choose"]
self.build_solutions = kwargs["operators"]["build_solutions"]
self.fitness = kwargs["operators"]["fitness"]
self.assign_credits = kwargs["operators"]["assign_credits"]
self.analyzers = kwargs.get("analyzers", [])
self.USE_CREDIT_INHERITANCE = kwargs.get("use_credit_inheritance", False)
self.HALL_OF_FAME_SIZE = kwargs.get("hall_of_fame_size", 0)
self._result = None
self.stat = tools.Statistics(key=lambda x: x.fitness)
self.stat.register("best", rounddec(numpy.max))
self.stat.register("min", rounddec(numpy.min))
self.stat.register("avg", rounddec(numpy.average))
self.stat.register("std", rounddec(numpy.std))
self.logbook = tools.Logbook()
self.kwargs['logbook'] = self.logbook
## TODO: make processing of population consisting of 1 element uniform
## generate initial population
self.pops = {s: self._generate_k(s.initialize(self.kwargs, s.pop_size)) if not s.fixed
else self._generate_k([s.representative_individual])
for s in self.SPECIES}
## make a copy for logging. TODO: remake it with logbook later.
self.initial_pops = {s.name: deepcopy(pop) for s, pop in self.pops.items()}
## checking correctness
for s, pop in self.pops.items():
sm = sum(p.k for p in pop)
assert sm == self.INTERACT_INDIVIDUALS_COUNT, \
"For specie {0} count doesn't equal to {1}. Actual value {2}".format(s, self.INTERACT_INDIVIDUALS_COUNT, sm)
print("Initialization finished")
self.hall = HallOfFame(self.HALL_OF_FAME_SIZE)
self.kwargs['gen'] = 0
pass
def __call__(self):
for gen in range(self.GENERATIONS):
self.gen()
print("Offsprings have been generated")
pass
return self.result()
def gen(self):
kwargs = self.kwargs
kwargs['gen'] = kwargs['gen'] + 1
print("Gen: " + str(kwargs['gen']))
solutions = self.build_solutions(self.pops, self.INTERACT_INDIVIDUALS_COUNT)
print("Solutions have been built")
## estimate fitness
for sol in solutions:
sol.fitness = self.fitness(kwargs, sol)
print("Fitness have been evaluated")
for s, pop in self.pops.items():
for p in pop:
p.fitness = -1000000000.0
## assign id, calculate credits and save it
i = 0
for s, pop in self.pops.items():
for p in pop:
p.id = i
i += 1
ind_maps = {p.id: p for s, pop in self.pops.items() for p in pop}
ind_to_credits = self.assign_credits(kwargs, solutions)
for ind_id, credit in ind_to_credits.items():
## assign credit to every individual
ind_maps[ind_id].fitness = credit
assert all([sum(p.fitness for p in pop) != 0 for s, pop in self.pops.items()]), \
"Error. Some population has individuals only with zero fitness"
print("Credit have been estimated")
solsstat_dict = dict(list(self.stat.compile(solutions).items()) + list(self.solstat(solutions).items()))
solsstat_dict["fitnesses"] = [sol.fitness for sol in solutions]
popsstat_dict = {s.name: dict(list(self.stat.compile(pop).items()) + list(s.stat(pop).items())) for s, pop in self.pops.items()}
for s, pop in self.pops.items():
popsstat_dict[s.name]["fitnesses"] = [p.fitness for p in pop]
if self.hall.maxsize > 0:
self.hall.update(solutions)
## TODO: this should be reconsidered
lsols = len(solutions)
solutions = list(self.hall) + solutions
solutions = solutions[0:lsols]
self.logbook.record(gen=kwargs['gen'],
popsstat=(popsstat_dict,),
solsstat=(solsstat_dict,))
for an in self.analyzers:
an(kwargs, solutions, self.pops)
print("hall: " + str(list(map(lambda x: x.fitness, self.hall))))
## select best solution as a result
## TODO: remake it in a more generic way
## TODO: add archive and corresponding updating and mixing
#best = max(solutions, key=lambda x: x.fitness)
## take the best
# best = hall[0] if hall.maxsize > 0 else max(solutions, key=lambda x: x.fitness)
self.best = self.hall[0] if self.hall.maxsize > 0 else max(solutions, key=lambda x: x.fitness)
## produce offsprings
items = [(s, pop) for s, pop in self.pops.items() if not s.fixed]
for s, pop in items:
if s.fixed:
continue
offspring = s.select(kwargs, pop)
offspring = list(map(deepcopy, offspring))
## apply mixin elite ones from the hall
if self.hall.maxsize > 0:
elite = [sol[s.name] for sol in self.hall]
offspring = elite + offspring
offspring = offspring[0:len(pop)]
# Apply crossover and mutation on the offspring
for child1, child2 in zip(offspring[::2], offspring[1::2]):
if random.random() < s.cxb:
c1 = child1.fitness
c2 = child2.fitness
s.mate(kwargs, child1, child2)
## TODO: make credit inheritance here
## TODO: toolbox.inherit_credit(pop, child1, child2)
## TODO: perhaps, this operation should be done after all crossovers in the pop
## default implementation
## ?
child1.fitness = (c1 + c2) / 2.0
child2.fitness = (c1 + c2) / 2.0
pass
for mutant in offspring:
if random.random() < s.mb:
s.mutate(kwargs, mutant)
pass
self.pops[s] = offspring
pass
## assign credits for every individuals of all pops
for s, pop in self.pops.items():
self._credit_to_k(pop)
for s, pop in self.pops.items():
for ind in pop:
if not self.USE_CREDIT_INHERITANCE:
del ind.fitness
del ind.id
pass
def result(self):
return self.best, self.pops, self.logbook, self.initial_pops
def _generate_k(self, pop):
base_k = int(self.INTERACT_INDIVIDUALS_COUNT / len(pop))
free_slots = self.INTERACT_INDIVIDUALS_COUNT % len(pop)
for ind in pop:
ind.k = base_k
for slot in range(free_slots):
i = random.randint(0, len(pop) - 1)
pop[i].k += 1
return pop
def _credit_to_k(self, pop):
norma = self.INTERACT_INDIVIDUALS_COUNT / sum(el.fitness for el in pop)
for c in pop:
c.k = int(c.fitness * norma)
left_part = self.INTERACT_INDIVIDUALS_COUNT - sum(c.k for c in pop)
sorted_pop = sorted(pop, key=lambda x: x.fitness, reverse=True)
for i in range(left_part):
sorted_pop[i].k += 1
return pop
pass
def _genetic_algo(model, initial_pop, exp_ess, base_biomass, toolbox, GA_param,
distance, processes, outputs):
index = [i for i in initial_pop.index]
print('made it to the GA')
print(outputs)
# Initial population evaluation
pop = toolbox.population_guess()
if 'history_name' in outputs.keys():
history.update(pop)
print('I will evaluate the initial population')
fits = toolbox.map(toolbox.evaluate, pop, initial_pop, model, base_biomass,
exp_ess, distance, processes)
print('Fitnesses that will be attributed to individuals %s' % (fits, ))
for ind, fit in zip(pop, fits):
ind.fitness.values = fit
print('Finished evaluating initial population')
# Record outputs
if 'hall_of_fame_name' in outputs.keys():
# Hall of fame
hof = HallOfFame(maxsize=outputs.get('hall_of_fame_size'))
hof.update(pop)
if 'logbook_name' in outputs.keys():
multi_level = pd.MultiIndex(
levels=[['Fitness', 'Size'],
['Mean', 'Std', 'Max', 'Min', 'Sizes']],
labels=[[
0,
0,
0,
0,
1,
], [0, 1, 2, 3, 4]])
data_logbook = []
#valid_logbook = []
# The GA itself
for g in range(GA_param.get('NGEN')):
print('Starting generation %s' % (g, ))
offspring = toolbox.select(pop, k=len(pop))
offspring = list(map(toolbox.clone, offspring))
for child1, child2 in zip(offspring[::2], offspring[1::2]):
# Keep the parents for comparison
parent1 = list(map(toolbox.clone, [child1]))[0]
parent2 = list(map(toolbox.clone, [child2]))[0]
rand_nb = random.random()
if rand_nb < GA_param.get('CXPB'):
print('doing a crossover')
if child1 == child2:
print('generating a new individual')
l = toolbox.generate_new_individual(
initial_pop.index, model, base_biomass)
child2 = creator.Individual(l)
if child2 == parent2:
print('generating new individual didnt work')
toolbox.mate(child1, child2)
# Assess the efficiency of crossover
if child1 == parent1 or child1 == parent2:
print('crossover yielded identical individuals')
l = toolbox.generate_new_individual(
initial_pop.index, model, base_biomass)
child2 = creator.Individual(l)
toolbox.mate(child1, child2)
else:
print('crossover actually worked')
del child1.fitness.values
del child2.fitness.values
for mutant in offspring:
rand_nb = random.random()
if rand_nb < GA_param.get('MUTPB'):
toolbox.mutate(mutant)
del mutant.fitness.values
# The individuals that have been crossed or mutated will have an invalid fitness
# and will need to be r-evaluated (saves on computing time)
invalid_ind = [ind for ind in offspring if not ind.fitness.valid]
print('I will evaluate %s invalid individuals, wish me luck!' %
(len(invalid_ind), ))
fitnesses = toolbox.map(toolbox.evaluate, invalid_ind, initial_pop,
model, base_biomass, exp_ess, distance,
processes)
print(fitnesses)
for ind, fit in zip(invalid_ind, fitnesses):
print(sum(ind))
ind.fitness.values = fit
print(" Evaluated %s individuals" % (len(invalid_ind), ))
pop[:] = offspring
# Save stats of the evolution
print('Trying to compile stats')
if 'logbook_name' in outputs.keys():
current_stats = record_stats(pop, g, multi_level)
# Manually update the logbook
data_logbook.append(current_stats)
if 'hall_of_fame_name' in outputs.keys():
hof.update(pop)
if 'history_name' in outputs.keys():
history.update(pop)
# Outputs
# Saving the logbook
if 'logbook_name' in outputs.keys():
final_logbook = pd.concat(data_logbook)
final_logbook.to_csv(outputs.get('logbook_name'))
# Get the best individuals from the hall of fame (metabolites)
if 'hall_of_fame_name' in outputs.keys():
hof_df = get_ind_composition(hof, index)
hof_df.to_csv(outputs.get('hall_of_fame_name'))
class PBIL(object):
train_scalars = {
'fitness_mean':
lambda population, last, overall: np.mean(
[x.fitness for x in population]),
'fitness_last':
lambda population, last, overall: last.fitness,
'fitness_overall':
lambda population, last, overall: overall.fitness
}
def __init__(self,
resources_path,
train_data,
lr=0.7,
selection_share=0.5,
n_generations=200,
n_individuals=75,
log_path=None):
"""
Initializes a new instance of PBIL ensemble learning classifier.
All PBIL hyper-parameters default to the values presented in the paper
Cagnini, Henry E.L., Freitas, Alex A., Barros, Rodrigo C.
An Evolutionary Algorithm for Learning Interpretable Ensembles of Classifiers.
Brazilian Conference on Intelligent Systems. 2020.
:param resources_path: Path to folder where at least two files must exist: classifiers.json and variables.json
:type resources_path: str
:param train_data: Training data as an object from the python weka wrapper library
:type train_data: weka.core.dataset.Instances
:param lr: Learning rate
:type lr: float
:param selection_share: How many individuals from general population with best fitness will update
graphical models' probabilities.
:type selection_share: float
:param n_generations: Number of generations to run PBIL
:type n_generations: int
:param n_individuals: Number of individuals (solutions) to use
:type n_individuals: int
:param log_path: Optional: path to where metadata from this run should be stored.
:type log_path: str
"""
self.lr = lr # type: float
self.selection_share = selection_share # type: float
self.n_generations = n_generations # type: int
self.n_individuals = n_individuals # type: int
clfs = [x[0] for x in inspect.getmembers(generation, inspect.isclass)]
classifier_names = [
x for x in clfs
if ClassifierWrapper in eval('generation.%s' % x).__bases__
]
self.classifier_names = classifier_names # type: list
self.variables = json.load(
open(os.path.join(resources_path, 'variables.json'),
'r')) # type: dict
self.classifier_data = json.load(
open(os.path.join(resources_path, 'classifiers.json'),
'r')) # type: dict
self.train_data = train_data # type: Instances
self.n_classes = len(self.train_data.class_attribute.values)
self.evaluator = EDAEvaluator(n_folds=5, train_data=self.train_data)
self.n_generation = 0
self._hof = None
scalars = copy.deepcopy(self.train_scalars)
if log_path is not None:
self.logger = PBILLogger(logdir_path=log_path,
histogram_names=['fitness'],
scalars=scalars,
text_names=['last', 'overall'])
self.logger.log_probabilities(
variables=self.variables) # register first probabilities
else:
self.logger = None
def sample_and_evaluate(self, seed, n_individuals):
"""
Samples new individuals from graphical model.
:param seed: seed used to partition training set at every generation. The (sub)sets will be constant throughout
all the evolutionary process, allowing a direct comparison between individuals from different generations.
:type seed: int
:param n_individuals: Number of individuals to sample.
:type n_individuals: int
:return: the recently sampled population
:rtype: list
"""
len_hall = len(self._hof)
if len_hall == 0:
parameters = {k: [] for k in self.classifier_names}
parameters['Aggregator'] = []
ilogs = []
else:
parameters = {
k: [x.options[k] for x in self._hof]
for k in self.classifier_names
}
parameters['Aggregator'] = [
x.options['Aggregator'] for x in self._hof
]
ilogs = [x.log for x in self._hof]
self._hof.clear()
for j in range(n_individuals):
ilog = dict()
for classifier_name in self.classifier_names:
ilog[classifier_name] = np.random.choice(
a=self.variables[classifier_name]['params']['a'],
p=self.variables[classifier_name]['params']['p'])
if ilog[classifier_name]: # whether to include this classifier in the ensemble
options, cclog = eval(classifier_name).sample_options(
variables=self.variables,
classifiers=self.classifier_data)
ilog.update(cclog)
parameters[classifier_name] += [options]
else:
parameters[classifier_name].append([])
ilog['Aggregator'] = np.random.choice(
a=self.variables['Aggregator']['params']['a'],
p=self.variables['Aggregator']['params']['p'])
agg_options, alog = eval(
ilog['Aggregator']).sample_options(variables=self.variables)
ilog.update(alog)
parameters['Aggregator'] += [[ilog['Aggregator']] + agg_options]
ilogs += [ilog]
train_aucs = self.evaluator.get_unweighted_aucs(seed=seed,
parameters=parameters)
# hall of fame is put in the front
for i in range(0, len_hall):
local_options = {
k: parameters[k][i]
for k in self.classifier_names
}
local_options['Aggregator'] = parameters['Aggregator'][i]
self._hof.insert(
Skeleton(seed=seed,
log=ilogs[i],
options=local_options,
fitness=train_aucs[i]))
population = []
for i in range(len_hall, n_individuals + len_hall):
local_options = {
k: parameters[k][i]
for k in self.classifier_names
}
local_options['Aggregator'] = parameters['Aggregator'][i]
population += [
Skeleton(seed=seed,
log=ilogs[i],
options=local_options,
fitness=train_aucs[i])
]
return population
def update(self, population):
"""
Updates graphical model probabilities based on the fittest population.
:param population: All population from a given generation.
:type population: list
"""
if self.logger is not None:
self.logger.log_probabilities(variables=self.variables)
self.logger.log_population(population=population,
halloffame=self._hof)
# selects fittest individuals
_sorted = sorted(zip(population, [ind.fitness for ind in population]),
key=lambda x: x[1],
reverse=True)
population, fitnesses = zip(*_sorted)
fittest = population[:int(len(population) * self.selection_share)]
observations = pd.DataFrame([fit.log for fit in fittest])
# update classifiers probabilities
for variable_name, variable_data in self.variables.items():
self.variables[variable_name] = process_update(
ptype=variable_data['ptype'],
variable_name=variable_name,
variable_data=variable_data,
observations=observations,
lr=self.lr,
n_generations=self.n_generations)
self.n_generation += 1
def run(self, seed):
"""
Trains this classifier.
:param seed: seed used to partition training set at every generation. The (sub)sets will be constant throughout
all the evolutionary process, allowing a direct comparison between individuals from different generations.
:type seed: int
:rtype: tuple
:return: a tuple containing two individuals.Individual objects, where the first individual is the best solution
(according to fitness) found throughout all the evolutionary process, and the second individual the best solution
from the last generation.
"""
# Statistics computation
stats = tools.Statistics(lambda ind: ind.fitness)
for stat_name, stat_func in PBILLogger.population_operators:
stats.register(stat_name, stat_func)
# An object that keeps track of the best individual found so far.
self._hof = HallOfFame(maxsize=1) # type: HallOfFame
best_last, logbook = self.__run__(seed=seed,
ngen=self.n_generations,
stats=stats,
verbose=True)
best_overall = self._hof[0] # type: Individual
self._hof = None
gc.collect()
return best_overall, best_last
def __run__(self, seed, ngen, stats=None, verbose=__debug__):
"""
Do not use this method.
"""
logbook = tools.Logbook()
logbook.header = ['gen', 'nevals'] + (stats.fields if stats else [])
early = EarlyStop()
population = []
for gen in range(ngen):
# early stop
if early.is_stopping():
break
# Generate a new population, already evaluated; re-evaluates halloffame with new seed
population = self.sample_and_evaluate(
seed=seed, n_individuals=self.n_individuals)
self._hof.update(population)
# Update the strategy with the evaluated individuals
self.update(population=population)
record = stats.compile(population) if stats is not None else {}
logbook.record(gen=gen, nevals=len(population), **record)
if verbose:
print(logbook.stream)
early.update(halloffame=self._hof, gen=gen)
fitnesses = [ind.fitness for ind in population]
best_skeleton = population[int(np.argmax(fitnesses))] # type: Skeleton
best_last = Individual(seed=seed,
log=best_skeleton.log,
options=best_skeleton.options,
train_data=self.train_data)
skts = [self._hof[i] for i in range(len(self._hof))]
self._hof.clear()
for i in range(len(skts)):
ind = Individual(seed=seed,
log=skts[i].log,
options=skts[i].options,
train_data=self.train_data)
self._hof.insert(ind)
return best_last, logbook
class NSGA2StatisticsCallback(object):
'''
Helper class for tracking statistics about the progression of the
optimization
'''
def __init__(self,
weights=(),
nr_of_generations=None,
crossover_rate=None,
mutation_rate=None,
pop_size=None,
caching=False):
'''
:param weights: the weights on the outcomes
:param nr_of_generations: the number of generations
:param crossover_rate: the crossover rate
:param mutation_rate: the mutation rate
:param caching: parameter controling wether a list of tried solutions
should be kept
'''
self.stats = []
if len(weights)>1:
self.hall_of_fame = ParetoFront(similar=compare)
else:
self.hall_of_fame = HallOfFame(pop_size)
if caching:
self.tried_solutions = TriedSolutions()
self.change = []
self.weights = weights
self.nr_of_generations = nr_of_generations
self.crossover_rate = crossover_rate
self.mutation_rate=mutation_rate
self.precision = "{0:.%if}" % 2
def __get_hof_in_array(self):
a = []
for entry in self.hall_of_fame:
a.append(entry.fitness.values)
return np.asarray(a)
def std(self, hof):
return np.std(hof, axis=0)
def mean(self, hof):
return np.mean(hof, axis=0)
def minima(self, hof):
return np.min(hof, axis=0)
def maxima(self, hof):
return np.max(hof, axis=0)
def log_stats(self, gen):
functions = {"minima":self.minima,
"maxima":self.maxima,
"std":self.std,
"mean":self.mean,}
kargs = {}
hof = self.__get_hof_in_array()
line = " ".join("{%s:<8}" % name for name in sorted(functions.keys()))
for name in sorted(functions.keys()):
function = functions[name]
kargs[name] = "[%s]" % ", ".join(map(self.precision.format, function(hof)))
line = line.format(**kargs)
line = "generation %s: " %gen + line
ema_logging.info(line)
def __call__(self, population):
self.change.append(self.hall_of_fame.update(population))
for entry in population:
self.stats.append(entry.fitness.values)
for entry in population:
try:
self.tried_solutions[entry] = entry
except AttributeError:
break
def init_population(self):
self.pop = self.toolbox.population(n=self.indi)
self.hof = HallOfFame(5, similar=np.array_equal)
class GeneticAlgorithmOptimizer(Optimizer):
"""
Implements evolutionary algorithm
:param ~l2l.utils.trajectory.Trajectory traj: Use this trajectory to store the parameters of the specific runs.
The parameters should be initialized based on the values in `parameters`
:param optimizee_create_individual: Function that creates a new individual
:param optimizee_fitness_weights: Fitness weights. The fitness returned by the Optimizee is multiplied by these
values (one for each element of the fitness vector)
:param parameters: Instance of :func:`~collections.namedtuple` :class:`.GeneticAlgorithmOptimizer` containing the parameters
needed by the Optimizer
"""
def __init__(self, traj,
optimizee_create_individual,
optimizee_fitness_weights,
parameters,
optimizee_bounding_func=None,
percent_hall_of_fame=0.2,
percent_elite=0.4,):
super().__init__(traj,
optimizee_create_individual=optimizee_create_individual,
optimizee_fitness_weights=optimizee_fitness_weights,
parameters=parameters, optimizee_bounding_func=optimizee_bounding_func)
self.optimizee_bounding_func = optimizee_bounding_func
__, self.optimizee_individual_dict_spec = \
dict_to_list(optimizee_create_individual(), get_dict_spec=True)
self.optimizee_create_individual = optimizee_create_individual
# if not len(traj.individuals):
popsize = parameters.popsize
traj.f_add_parameter('seed', parameters.seed, comment='Seed for RNG')
traj.f_add_parameter('popsize', popsize, comment='Population size') # 185
traj.f_add_parameter('CXPB', parameters.CXPB, comment='Crossover term')
traj.f_add_parameter('MUTPB', parameters.MUTPB, comment='Mutation probability')
traj.f_add_parameter('n_iteration', parameters.NGEN, comment='Number of generations')
traj.f_add_parameter('indpb', parameters.indpb, comment='Mutation parameter')
traj.f_add_parameter('tournsize', parameters.tournsize, comment='Selection parameter')
# ------- Create and register functions with DEAP ------- #
# delay_rate, slope, std_err, max_fraction_active
creator.create("FitnessMax", base.Fitness, weights=self.optimizee_fitness_weights)
creator.create("Individual", list, fitness=creator.FitnessMax)
toolbox = base.Toolbox()
# Structure initializers
toolbox.register("individual", tools.initIterate, creator.Individual,
lambda: dict_to_list(optimizee_create_individual()))
toolbox.register("population", tools.initRepeat, list, toolbox.individual)
# Operator registering
# This complex piece of code is only necessary because we're using the
# DEAP framework and would like to decorate the DEAP mutation operator
def bounding_decorator(func):
def bounding_wrapper(*args, **kwargs):
if self.optimizee_bounding_func is None:
return func(*args, **kwargs)
else:
# Deap Functions modify individuals in-place, Hence we must do the same
result_individuals_deap = func(*args, **kwargs)
result_individuals = [list_to_dict(x, self.optimizee_individual_dict_spec)
for x in result_individuals_deap]
bounded_individuals = [self.optimizee_bounding_func(x) for x in result_individuals]
for i, deap_indiv in enumerate(result_individuals_deap):
deap_indiv[:] = dict_to_list(bounded_individuals[i])
# print("Bounded Individual: {}".format(bounded_individuals))
return result_individuals_deap
return bounding_wrapper
toolbox.register("mate", tools.cxBlend, alpha=parameters.matepar)
toolbox.decorate("mate", bounding_decorator)
toolbox.register("mutate", tools.mutGaussian, mu=0, sigma=parameters.mutpar, indpb=traj.indpb)
toolbox.decorate("mutate", bounding_decorator)
toolbox.register("select", tools.selTournament, tournsize=traj.tournsize)
# ------- Initialize Population and Trajectory -------- #
# NOTE: The Individual object implements the list interface.
self.pop = toolbox.population(n=traj.popsize)
# traj.individuals.clear()
self.n_hof = int(max(percent_hall_of_fame * traj.popsize, 2))
self.n_bobs = int(max(1, percent_elite * self.n_hof))
self.hall_of_fame = HallOfFame(self.n_hof)
if len(traj.individuals):
self._load_last_trajectories(traj)
self.hall_of_fame.update(self.pop)
bob_inds = tools.selBest(self.hall_of_fame, self.n_bobs)
for hof_ind in bob_inds:
sind = str_ind(
list_to_dict(hof_ind, self.optimizee_individual_dict_spec)
)
logger.info("HOF individual is %s, \n%s" % (
sind, hof_ind.fitness.values))
# print("HOF individual is %s, %s" % (
# hof_ind, hof_ind.fitness.values))
print("Starting BoB individual is %s, \n%s" % (
sind, hof_ind.fitness.values))
self._delete_initial_fitnesses()
self.eval_pop_inds = [ind for ind in self.pop if not ind.fitness.valid]
self.eval_pop = [list_to_dict(ind, self.optimizee_individual_dict_spec)
for ind in self.eval_pop_inds]
if len(traj.individuals):
g = np.max(list(traj.individuals.keys()))
else:
g = 0
self.g = g # the current generation
self.toolbox = toolbox # the DEAP toolbox
self._expand_trajectory(traj)
x = traj
def _delete_initial_fitnesses(self):
for ind in self.pop:
del ind.fitness.values
def _load_last_trajectories(self, traj):
if not len(traj.individuals):
return
g = np.max(list(traj.individuals.keys()))
res = traj.results._data['all_results']._data[g]
for idx, ind in enumerate(traj.individuals[g]):
for k_idx, k in enumerate(sorted(ind.keys)):
sk = k.split('.')[-1]
v = getattr(ind, sk)
self.pop[idx][k_idx] = v
self.pop[idx].fitness.values = res[idx][1]
x = self.pop[idx]
# del self.pop[idx].fitness.values
def post_process(self, traj, fitnesses_results):
"""
See :meth:`~l2l.optimizers.optimizer.Optimizer.post_process`
"""
def to_fit(ind):
return np.dot(ind.fitness.values, ind.fitness.weights)
def spawn():
x = self.optimizee_create_individual()
return dict_to_list(self.optimizee_bounding_func(x))
CXPB, MUTPB, NGEN = traj.CXPB, traj.MUTPB, traj.n_iteration
logger.info(" Evaluating %i individuals" % len(fitnesses_results))
print(" Evaluating %i individuals" % len(fitnesses_results))
#******************************************************************
# Storing run-information in the trajectory
# Reading fitnesses and performing distribution update
#******************************************************************
# print("self.g = {}".format(self.g))
# print("len(self.eval_pop_inds) = {}".format(len(self.eval_pop_inds)))
# print("len(self.eval_pop) = {}".format(len(self.eval_pop)))
# print("len(fitnesses_results) = {}".format(len(fitnesses_results)))
for run_index, fitness in fitnesses_results:
# We need to convert the current run index into an ind_idx
# (index of individual within one generation)
traj.v_idx = run_index
ind_index = traj.par.ind_idx
traj.f_add_result('$set.$.individual', self.eval_pop[ind_index])
traj.f_add_result('$set.$.fitness', fitness)
# Use the ind_idx to update the fitness
individual = self.eval_pop_inds[ind_index]
individual.fitness.values = fitness
traj.v_idx = -1 # set the trajectory back to default
logger.info("-- End of generation {} --".format(self.g))
print("-- End of generation {} --".format(self.g))
# best_inds = tools.selBest(self.eval_pop_inds, 2)
# for best_ind in best_inds:
# print("Best individual is %s, %s" % (
# list_to_dict(best_ind, self.optimizee_individual_dict_spec),
# best_ind.fitness.values))
# add the bestest individuals this generation to HoF
self.hall_of_fame.update(self.eval_pop_inds)
logger.info("-- Hall of fame --")
# n_bobs = self.n_bobs
# bob_inds = tools.selBest(self.hall_of_fame, n_bobs)
n_bobs = self.n_hof
bob_inds = tools.selBest(self.hall_of_fame, n_bobs)
for hof_ind in self.hall_of_fame:
sind = str_ind(
list_to_dict(hof_ind, self.optimizee_individual_dict_spec)
)
logger.info("BoB individual is %s,\n %s" % (
sind, hof_ind.fitness.values))
# print("HOF individual is %s, %s" % (
# hof_ind, hof_ind.fitness.values))
print("BoB individual is %s,\n %s" % (
sind, hof_ind.fitness.values))
#bob_inds = list(map(self.toolbox.clone, bob_inds))
bob_inds = list(map(self.toolbox.clone, self.hall_of_fame))
# ------- Create the next generation by crossover and mutation -------- #
if self.g < NGEN - 1: # not necessary for the last generation
# Select the next generation individuals
# Tournament of population - a list of "pointers"
offspring = self.toolbox.select(self.pop, len(self.pop))
# Clone the selected individuals
offspring = list(map(self.toolbox.clone, offspring))
#sorts small to big
#switch worst-good with best of best
#ascending (worst to best)
offsp_ids = np.argsort([to_fit(o) for o in offspring])
#descending (best to worst)
bob_ids = np.argsort([to_fit(o) for o in bob_inds])[::-1]
max_score = to_fit(bob_inds[bob_ids[0]])
min_score = 0.5 * max_score
for i in range(n_bobs):
off_i = int(offsp_ids[i])
bob_i = int(bob_ids[i])
off_f = to_fit(offspring[off_i])
bob_f = to_fit(bob_inds[bob_i])
if bob_f > off_f:
logger.info("Inserting BoB {} to population".format(i+1))
offspring[off_i][:] = bob_inds[bob_i]
# Apply crossover and mutation on the offspring
for child1, child2 in zip(offspring[::2], offspring[1::2]):
f1, f2 = to_fit(child1), to_fit(child2)
if random.random() < CXPB:
#if both parents are really unfit, replace them with a new couple
if f1 <= min_score and f2 <= min_score:
logger.info("Both parents had a low score")
child1[:] = spawn()
child2[:] = spawn()
else:
self.toolbox.mate(child1, child2)
del child1.fitness.values
del child2.fitness.values
for mutant in offspring[:]:
if random.random() < MUTPB:
# f = to_fit(mutant) if mutant.fitness.valid else None
# print("f = {}".format(f))
# if this was an unfit individual, replace with a "foreigner"
# if f is not None and f <= min_score:
# logger.info("Mutant had a really low score")
# mutant[:] = spawn()
# else:
self.toolbox.mutate(mutant)
del mutant.fitness.values
if len(set(map(tuple, offspring))) < len(offspring):
logger.info("Mutating more")
for i, o1 in enumerate(offspring[:-1]):
for o2 in offspring[i+1:]:
if tuple(np.round(o1, decimals=4)) == tuple(np.round(o2, decimals=4)):
if random.random() < 0.8:
self.toolbox.mutate(o2)
#o2[:] = spawn()
del o2.fitness.values
# off_ids = np.random.choice(len(offspring), size=n_bobs, replace=False)
# for i in range(n_bobs):
# # off_i = int(offsp_ids[i])
# off_i = int(off_ids[i])
# bob_i = int(bob_ids[i])
# logger.info("Inserting BoB {} to population".format(i+1))
# offspring[off_i][:] = bob_inds[bob_i]
# del offspring[off_i].fitness.values
# The population is entirely replaced by the offspring
self.pop[:] = offspring
self.eval_pop_inds[:] = [ind for ind in self.pop if not ind.fitness.valid]
self.eval_pop[:] = [list_to_dict(ind, self.optimizee_individual_dict_spec)
for ind in self.eval_pop_inds]
# print("self.g = {}".format(self.g))
# print("len(self.eval_pop_inds) = {}".format(len(self.eval_pop_inds)))
# print("len(self.eval_pop) = {}".format(len(self.eval_pop)))
self.g += 1 # Update generation counter
if len(self.eval_pop) == 0 and self.g < (NGEN - 1):
raise Exception("No more mutants to evaluate where generated. "
"Increasing population size may help.")
self._expand_trajectory(traj)
# print("self.g = {}".format(self.g))
# print("len(self.eval_pop_inds) = {}".format(len(self.eval_pop_inds)))
# print("len(self.eval_pop) = {}".format(len(self.eval_pop)))
def end(self, traj):
"""
See :meth:`~l2l.optimizers.optimizer.Optimizer.end`
"""
# ------------ Finished all runs and print result --------------- #
logger.info("-- End of (successful) evolution --")
best_inds = tools.selBest(self.pop, 10)
for best_ind in best_inds:
logger.info("Best individual is %s, %s" % (best_ind, best_ind.fitness.values))
logger.info("-- Hall of fame --")
for hof_ind in self.hall_of_fame:
logger.info("HOF individual is %s, %s" % (hof_ind, hof_ind.fitness.values))
def new_hall_of_fame(self) -> HallOfFame:
return HallOfFame(maxsize=50)
def __init__(self, pset, pop=500, gen=20, mutate_prob=0.5, mate_prob=0.8, hall=1, re_hall=1,
re_Tree=None, initial_min=None, initial_max=3, max_value=5,
scoring=(r2_score,), score_pen=(1,), filter_warning=True, cv=1,
add_coef=True, inter_add=True, inner_add=False, vector_add=False, out_add=False, flat_add=False,
cal_dim=False, dim_type=None, fuzzy=False, n_jobs=1, batch_size=40,
random_state=None, stats=None, verbose=True, migrate_prob=0,
tq=True, store=False, personal_map=False, stop_condition=None, details=False, classification=False,
score_object="y", sub_mu_max=1, limit_type="h_bgp", batch_para=False):
"""
Parameters
----------
pset:SymbolSet
the feature x and target y and others should have been added.
pop: int
number of population.
gen: int
number of generation.
mutate_prob:float
probability of mutate.
mate_prob:float
probability of mate(crossover).
initial_max:int
max initial size of expression when first producing.
initial_min : None,int
min initial size of expression when first producing.
max_value:int
max size of expression.
limit_type: "height" or "length",","h_bgp"
limitation type for max_value, but don't affect initial_max, initial_min.
hall:int,>=1
number of HallOfFame (elite) to maintain.
re_hall:None or int>=2
Notes: only valid when hall
number of HallOfFame to add to next generation.
re_Tree: int
number of new features to add to next generation.
0 is false to add.
personal_map:bool or "auto"
"auto" is using 'premap' and with auto refresh the 'premap' with individual.\n
True is just using constant 'premap'.\n
False is just use the prob of terminals.
scoring: list of Callable, default is [sklearn.metrics.r2_score,]
See Also ``sklearn.metrics``
score_pen: tuple of 1, -1 or float but 0.
>0 : max problem, best is positive, worse -np.inf.
<0 : min problem, best is negative, worse np.inf.
Notes:
if multiply score method, the scores must be turn to same dimension in prepossessing
or weight by score_pen. Because the all the selection are stand on the mean(w_i*score_i)
Examples::
scoring = [r2_score,]
score_pen= [1,]
cv:sklearn.model_selection._split._BaseKFold,int
the shuffler must be False,
default=1 means no cv.
filter_warning:bool
filter warning or not.
add_coef:bool
add coef in expression or not.
inter_add：bool
add intercept constant or not.
inner_add:bool
add inner coefficients or not.
out_add:bool
add out coefficients or not.
flat_add:bool
add flat coefficients or not.
n_jobs:int
default 1, advise 6.
batch_size:int
default 40, depend of machine.
random_state:int
None,int.
cal_dim:bool
escape the dim calculation.
dim_type:Dim or None or list of Dim
"coef": af(x)+b. a,b have dimension,f(x)'s dimension is not dnan. \n
"integer": af(x)+b. f(x) is with integer dimension. \n
[Dim1,Dim2]: f(x)'s dimension in list. \n
Dim: f(x) ~= Dim. (see fuzzy) \n
Dim: f(x) == Dim. \n
None: f(x) == pset.y_dim
fuzzy:bool
choose the dim with same base with dim_type, such as m,m^2,m^3.
stats:dict
details of logbook to show. \n
Map:\n
values
= {"max": np.max, "mean": np.mean, "min": np.mean, "std": np.std, "sum": np.sum}
keys
= {\n
"fitness": just see fitness[0], \n
"fitness_dim_max": max problem, see fitness with demand dim,\n
"fitness_dim_min": min problem, see fitness with demand dim,\n
"dim_is_target": demand dim,\n
"coef": dim is True, coef have dim, \n
"integer": dim is integer, \n
...
}
if stats is None, default is:
for cal_dim=True:
stats = {"fitness_dim_max": ("max",), "dim_is_target": ("sum",)}
for cal_dim=False
stats = {"fitness": ("max",)}
if self-definition, the key is func to get attribute of each ind.
Examples::
def func(ind):
return ind.fitness[0]
stats = {func: ("mean",), "dim_is_target": ("sum",)}
verbose:bool
print verbose logbook or not.
tq:bool
print progress bar or not.
store:bool or path
bool or path.
stop_condition:callable
stop condition on the best ind of hall, which return bool,the true means stop loop.
Examples::
def func(ind):
c = ind.fitness.values[0]>=0.90
return c
details:bool
return expr and predict_y or not.
classification: bool
classification or not.
score_object:
score by y or delta y (for implicit function).
"""
super(BaseLoop, self).__init__()
assert initial_max <= max_value, "the initial size of expression should less than max_value limitation"
if cal_dim:
assert all(
[isinstance(i, Dim) for i in pset.dim_ter_con.values()]), \
"all import dim of pset should be Dim object."
self.details = details
self.max_value = max_value
self.pop = pop
self.gen = gen
self.mutate_prob = mutate_prob
self.mate_prob = mate_prob
self.migrate_prob = migrate_prob
self.verbose = verbose
self.cal_dim = cal_dim
self.re_hall = re_hall
self.re_Tree = re_Tree
self.store = store
self.limit_type = limit_type
self.data_all = []
self.personal_map = personal_map
self.stop_condition = stop_condition
self.population = []
self.rand_state = None
self.random_state = random_state
self.sub_mu_max = sub_mu_max
self.population_next = []
self.cpset = CalculatePrecisionSet(pset, scoring=scoring, score_pen=score_pen,
filter_warning=filter_warning, cal_dim=cal_dim,
add_coef=add_coef, inter_add=inter_add, inner_add=inner_add,
vector_add=vector_add, out_add=out_add, flat_add=flat_add, cv=cv,
n_jobs=n_jobs, batch_size=batch_size, tq=tq,
fuzzy=fuzzy, dim_type=dim_type, details=details,
classification=classification, score_object=score_object,
batch_para=batch_para
)
Fitness_ = newclass.create("Fitness_", Fitness, weights=score_pen)
self.PTree = newclass.create("PTrees", SymbolTree, fitness=Fitness_)
# def produce
if initial_min is None:
initial_min = 2
self.register("genGrow", genGrow, pset=self.cpset, min_=initial_min, max_=initial_max + 1,
personal_map=self.personal_map)
self.register("genFull", genFull, pset=self.cpset, min_=initial_min, max_=initial_max + 1,
personal_map=self.personal_map)
self.register("genHalf", genGrow, pset=self.cpset, min_=initial_min, max_=initial_max + 1,
personal_map=self.personal_map)
self.register("gen_mu", genGrow, min_=1, max_=self.sub_mu_max + 1, personal_map=self.personal_map)
# def selection
self.register("select", selTournament, tournsize=2)
self.register("selKbestDim", selKbestDim,
dim_type=self.cpset.dim_type, fuzzy=self.cpset.fuzzy)
self.register("selBest", selBest)
self.register("mate", cxOnePoint)
# def mutate
self.register("mutate", mutUniform, expr=self.gen_mu, pset=self.cpset)
self.decorate("mate", staticLimit(key=operator.attrgetter(limit_type), max_value=self.max_value))
self.decorate("mutate", staticLimit(key=operator.attrgetter(limit_type), max_value=self.max_value))
if stats is None:
if cal_dim:
if score_pen[0] > 0:
stats = {"fitness_dim_max": ("max",), "dim_is_target": ("sum",)}
else:
stats = {"fitness_dim_min": ("min",), "dim_is_target": ("sum",)}
else:
if score_pen[0] > 0:
stats = {"fitness": ("max",)}
else:
stats = {"fitness": ("min",)}
self.stats = Statis_func(stats=stats)
logbook = Logbook()
logbook.header = ['gen'] + (self.stats.fields if self.stats else [])
self.logbook = logbook
if hall is None:
hall = 1
self.hall = HallOfFame(hall)
if re_hall is None:
self.re_hall = None
else:
if re_hall == 1 or re_hall == 0:
print("re_hall should more than 1")
re_hall = 2
assert re_hall >= hall, "re_hall should more than hall"
self.re_hall = HallOfFame(re_hall)
class GeneticAlgorithmOptimizer(Optimizer):
"""
Implements evolutionary algorithm
:param ~l2l.utils.trajectory.Trajectory traj: Use this trajectory to store the parameters of the specific runs.
The parameters should be initialized based on the values in `parameters`
:param optimizee_create_individual: Function that creates a new individual
:param optimizee_fitness_weights: Fitness weights. The fitness returned by the Optimizee is multiplied by these
values (one for each element of the fitness vector)
:param parameters: Instance of :func:`~collections.namedtuple` :class:`.GeneticAlgorithmOptimizer` containing the parameters
needed by the Optimizer
"""
def __init__(self,
traj,
optimizee_create_individual,
optimizee_fitness_weights,
parameters,
optimizee_bounding_func=None):
super().__init__(
traj,
optimizee_create_individual=optimizee_create_individual,
optimizee_fitness_weights=optimizee_fitness_weights,
parameters=parameters,
optimizee_bounding_func=optimizee_bounding_func)
self.optimizee_bounding_func = optimizee_bounding_func
__, self.optimizee_individual_dict_spec = dict_to_list(
optimizee_create_individual(), get_dict_spec=True)
traj.f_add_parameter('seed', parameters.seed, comment='Seed for RNG')
traj.f_add_parameter('popsize',
parameters.popsize,
comment='Population size') # 185
traj.f_add_parameter('CXPB', parameters.CXPB, comment='Crossover term')
traj.f_add_parameter('MUTPB',
parameters.MUTPB,
comment='Mutation probability')
traj.f_add_parameter('n_iteration',
parameters.NGEN,
comment='Number of generations')
traj.f_add_parameter('indpb',
parameters.indpb,
comment='Mutation parameter')
traj.f_add_parameter('tournsize',
parameters.tournsize,
comment='Selection parameter')
# ------- Create and register functions with DEAP ------- #
# delay_rate, slope, std_err, max_fraction_active
creator.create("FitnessMax",
base.Fitness,
weights=self.optimizee_fitness_weights)
creator.create("Individual", list, fitness=creator.FitnessMax)
toolbox = base.Toolbox()
# Structure initializers
toolbox.register("individual", tools.initIterate, creator.Individual,
lambda: dict_to_list(optimizee_create_individual()))
toolbox.register("population", tools.initRepeat, list,
toolbox.individual)
# Operator registering
# This complex piece of code is only necessary because we're using the
# DEAP framework and would like to decorate the DEAP mutation operator
def bounding_decorator(func):
def bounding_wrapper(*args, **kwargs):
if self.optimizee_bounding_func is None:
return func(*args, **kwargs)
else:
# Deap Functions modify individuals in-place, Hence we must do the same
result_individuals_deap = func(*args, **kwargs)
result_individuals = [
list_to_dict(x, self.optimizee_individual_dict_spec)
for x in result_individuals_deap
]
bounded_individuals = [
self.optimizee_bounding_func(x)
for x in result_individuals
]
for i, deap_indiv in enumerate(result_individuals_deap):
deap_indiv[:] = dict_to_list(bounded_individuals[i])
print("Bounded Individual: {}".format(bounded_individuals))
return result_individuals_deap
return bounding_wrapper
toolbox.register("mate", tools.cxBlend, alpha=parameters.matepar)
toolbox.decorate("mate", bounding_decorator)
toolbox.register("mutate",
tools.mutGaussian,
mu=0,
sigma=parameters.mutpar,
indpb=traj.indpb)
toolbox.decorate("mutate", bounding_decorator)
toolbox.register("select",
tools.selTournament,
tournsize=traj.tournsize)
# ------- Initialize Population and Trajectory -------- #
# NOTE: The Individual object implements the list interface.
self.pop = toolbox.population(n=traj.popsize)
self.eval_pop_inds = [ind for ind in self.pop if not ind.fitness.valid]
self.eval_pop = [
list_to_dict(ind, self.optimizee_individual_dict_spec)
for ind in self.eval_pop_inds
]
self.g = 0 # the current generation
self.toolbox = toolbox # the DEAP toolbox
self.hall_of_fame = HallOfFame(20)
self._expand_trajectory(traj)
def post_process(self, traj, fitnesses_results):
"""
See :meth:`~l2l.optimizers.optimizer.Optimizer.post_process`
"""
CXPB, MUTPB, NGEN = traj.CXPB, traj.MUTPB, traj.n_iteration
logger.info(" Evaluating %i individuals" % len(fitnesses_results))
#**************************************************************************************************************
# Storing run-information in the trajectory
# Reading fitnesses and performing distribution update
#**************************************************************************************************************
for run_index, fitness in fitnesses_results:
# We need to convert the current run index into an ind_idx
# (index of individual within one generation)
traj.v_idx = run_index
ind_index = traj.par.ind_idx
traj.f_add_result('$set.$.individual', self.eval_pop[ind_index])
traj.f_add_result('$set.$.fitness', fitness)
# Use the ind_idx to update the fitness
individual = self.eval_pop_inds[ind_index]
individual.fitness.values = fitness
traj.v_idx = -1 # set the trajectory back to default
logger.info("-- End of generation {} --".format(self.g))
best_inds = tools.selBest(self.eval_pop_inds, 2)
for best_ind in best_inds:
print("Best individual is %s, %s" %
(list_to_dict(best_ind, self.optimizee_individual_dict_spec),
best_ind.fitness.values))
self.hall_of_fame.update(self.eval_pop_inds)
logger.info("-- Hall of fame --")
for hof_ind in tools.selBest(self.hall_of_fame, 2):
logger.info(
"HOF individual is %s, %s" %
(list_to_dict(hof_ind, self.optimizee_individual_dict_spec),
hof_ind.fitness.values))
# ------- Create the next generation by crossover and mutation -------- #
if self.g < NGEN - 1: # not necessary for the last generation
# Select the next generation individuals
offspring = self.toolbox.select(self.pop, len(self.pop))
# Clone the selected individuals
offspring = list(map(self.toolbox.clone, offspring))
# Apply crossover and mutation on the offspring
for child1, child2 in zip(offspring[::2], offspring[1::2]):
if random.random() < CXPB:
self.toolbox.mate(child1, child2)
del child1.fitness.values
del child2.fitness.values
for mutant in offspring:
if random.random() < MUTPB:
self.toolbox.mutate(mutant)
del mutant.fitness.values
if len(set(map(tuple, offspring))) < len(offspring):
logger.info("Mutating more")
for i, o1 in enumerate(offspring[:-1]):
for o2 in offspring[i + 1:]:
if tuple(o1) == tuple(o2):
if random.random() < 0.8:
self.toolbox.mutate(o2)
del o2.fitness.values
# The population is entirely replaced by the offspring
self.pop[:] = offspring
self.eval_pop_inds = [
ind for ind in self.pop if not ind.fitness.valid
]
self.eval_pop = [
list_to_dict(ind, self.optimizee_individual_dict_spec)
for ind in self.eval_pop_inds
]
self.g += 1 # Update generation counter
self._expand_trajectory(traj)
def end(self):
"""
See :meth:`~l2l.optimizers.optimizer.Optimizer.end`
"""
# ------------ Finished all runs and print result --------------- #
logger.info("-- End of (successful) evolution --")
best_inds = tools.selBest(self.pop, 10)
for best_ind in best_inds:
logger.info("Best individual is %s, %s" %
(best_ind, best_ind.fitness.values))
logger.info("-- Hall of fame --")
for hof_ind in self.hall_of_fame:
logger.info("HOF individual is %s, %s" %
(hof_ind, hof_ind.fitness.values))
def mainPart(x_,
y_,
pset,
max_=5,
pop_n=100,
random_seed=2,
cxpb=0.8,
mutpb=0.1,
ngen=5,
tournsize=3,
max_value=10,
double=False,
score=None,
cal_dim=True,
target_dim=None,
inter_add=True,
iner_add=True,
random_add=False,
store=True):
"""
Parameters
----------
target_dim
max_
inter_add
iner_add
random_add
cal_dim
score
double
x_
y_
pset
pop_n
random_seed
cxpb
mutpb
ngen
tournsize
max_value
Returns
-------
"""
if score is None:
score = [r2_score, explained_variance_score]
if cal_dim:
assert all([isinstance(i, Dim) for i in pset.dim_list
]), "all import dim of pset should be Dim object"
random.seed(random_seed)
toolbox = Toolbox()
if isinstance(pset, ExpressionSet):
PTrees = ExpressionTree
Generate = genHalfAndHalf
mutate = mutNodeReplacement
mate = cxOnePoint
elif isinstance(pset, FixedSet):
PTrees = FixedTree
Generate = generate_index
mutate = mutUniForm_index
mate = partial(cxOnePoint_index, pset=pset)
else:
raise NotImplementedError("get wrong pset")
if double:
Fitness_ = creator.create("Fitness_", Fitness, weights=(1.0, 1.0))
else:
Fitness_ = creator.create("Fitness_", Fitness, weights=(1.0, ))
PTrees_ = creator.create("PTrees_",
PTrees,
fitness=Fitness_,
dim=dnan,
withdim=0)
toolbox.register("generate", Generate, pset=pset, min_=1, max_=max_)
toolbox.register("individual",
initIterate,
container=PTrees_,
generator=toolbox.generate)
toolbox.register('population',
initRepeat,
container=list,
func=toolbox.individual)
# def selection
toolbox.register("select_gs", selTournament, tournsize=tournsize)
toolbox.register("select_kbest_target_dim",
selKbestDim,
dim_type=target_dim,
fuzzy=True)
toolbox.register("select_kbest_dimless", selKbestDim, dim_type="integer")
toolbox.register("select_kbest", selKbestDim, dim_type='ignore')
# def mate
toolbox.register("mate", mate)
# def mutate
toolbox.register("mutate", mutate, pset=pset)
if isinstance(pset, ExpressionSet):
toolbox.decorate(
"mate",
staticLimit(key=operator.attrgetter("height"),
max_value=max_value))
toolbox.decorate(
"mutate",
staticLimit(key=operator.attrgetter("height"),
max_value=max_value))
# def elaluate
toolbox.register("evaluate",
calculatePrecision,
pset=pset,
x=x_,
y=y_,
scoring=score[0],
cal_dim=cal_dim,
inter_add=inter_add,
iner_add=iner_add,
random_add=random_add)
toolbox.register("evaluate2",
calculatePrecision,
pset=pset,
x=x_,
y=y_,
scoring=score[1],
cal_dim=cal_dim,
inter_add=inter_add,
iner_add=iner_add,
random_add=random_add)
toolbox.register("parallel",
parallelize,
n_jobs=1,
func=toolbox.evaluate,
respective=False)
toolbox.register("parallel2",
parallelize,
n_jobs=1,
func=toolbox.evaluate2,
respective=False)
pop = toolbox.population(n=pop_n)
haln = 10
hof = HallOfFame(haln)
stats1 = Statistics(lambda ind: ind.fitness.values[0]
if ind and ind.y_dim in target_dim else 0)
stats1.register("max", np.max)
stats2 = Statistics(lambda ind: ind.y_dim in target_dim if ind else 0)
stats2.register("countable_number", np.sum)
stats = MultiStatistics(score1=stats1, score2=stats2)
population, logbook = eaSimple(pop,
toolbox,
cxpb=cxpb,
mutpb=mutpb,
ngen=ngen,
stats=stats,
halloffame=hof,
pset=pset,
store=store)
return population, hof
def __init__(self,
traj,
optimizee_create_individual,
optimizee_fitness_weights,
parameters,
optimizee_bounding_func=None):
super().__init__(
traj,
optimizee_create_individual=optimizee_create_individual,
optimizee_fitness_weights=optimizee_fitness_weights,
parameters=parameters,
optimizee_bounding_func=optimizee_bounding_func)
self.optimizee_bounding_func = optimizee_bounding_func
__, self.optimizee_individual_dict_spec = dict_to_list(
optimizee_create_individual(), get_dict_spec=True)
traj.f_add_parameter('seed', parameters.seed, comment='Seed for RNG')
traj.f_add_parameter('popsize',
parameters.popsize,
comment='Population size') # 185
traj.f_add_parameter('CXPB', parameters.CXPB, comment='Crossover term')
traj.f_add_parameter('MUTPB',
parameters.MUTPB,
comment='Mutation probability')
traj.f_add_parameter('n_iteration',
parameters.NGEN,
comment='Number of generations')
traj.f_add_parameter('indpb',
parameters.indpb,
comment='Mutation parameter')
traj.f_add_parameter('tournsize',
parameters.tournsize,
comment='Selection parameter')
# ------- Create and register functions with DEAP ------- #
# delay_rate, slope, std_err, max_fraction_active
creator.create("FitnessMax",
base.Fitness,
weights=self.optimizee_fitness_weights)
creator.create("Individual", list, fitness=creator.FitnessMax)
toolbox = base.Toolbox()
# Structure initializers
toolbox.register("individual", tools.initIterate, creator.Individual,
lambda: dict_to_list(optimizee_create_individual()))
toolbox.register("population", tools.initRepeat, list,
toolbox.individual)
# Operator registering
# This complex piece of code is only necessary because we're using the
# DEAP framework and would like to decorate the DEAP mutation operator
def bounding_decorator(func):
def bounding_wrapper(*args, **kwargs):
if self.optimizee_bounding_func is None:
return func(*args, **kwargs)
else:
# Deap Functions modify individuals in-place, Hence we must do the same
result_individuals_deap = func(*args, **kwargs)
result_individuals = [
list_to_dict(x, self.optimizee_individual_dict_spec)
for x in result_individuals_deap
]
bounded_individuals = [
self.optimizee_bounding_func(x)
for x in result_individuals
]
for i, deap_indiv in enumerate(result_individuals_deap):
deap_indiv[:] = dict_to_list(bounded_individuals[i])
print("Bounded Individual: {}".format(bounded_individuals))
return result_individuals_deap
return bounding_wrapper
toolbox.register("mate", tools.cxBlend, alpha=parameters.matepar)
toolbox.decorate("mate", bounding_decorator)
toolbox.register("mutate",
tools.mutGaussian,
mu=0,
sigma=parameters.mutpar,
indpb=traj.indpb)
toolbox.decorate("mutate", bounding_decorator)
toolbox.register("select",
tools.selTournament,
tournsize=traj.tournsize)
# ------- Initialize Population and Trajectory -------- #
# NOTE: The Individual object implements the list interface.
self.pop = toolbox.population(n=traj.popsize)
self.eval_pop_inds = [ind for ind in self.pop if not ind.fitness.valid]
self.eval_pop = [
list_to_dict(ind, self.optimizee_individual_dict_spec)
for ind in self.eval_pop_inds
]
self.g = 0 # the current generation
self.toolbox = toolbox # the DEAP toolbox
self.hall_of_fame = HallOfFame(20)
self._expand_trajectory(traj)
def evolution(env: Environment, number_of_rays: int, ray_distribution: str,
angle_lower_bound: int, angle_upper_bound: int,
length_lower_bound: int, length_upper_bound: int,
no_of_reflective_segments: int, distance_limit: int,
length_limit: int, population_size: int,
number_of_generations: int, xover_prob: float,
mut_angle_prob: float, mut_length_prob: float,
shift_segment_prob: float, rotate_segment_prob: float,
resize_segment_prob: float):
# Initiating evolutionary algorithm
creator.create("Fitness", base.Fitness, weights=(1.0, ))
creator.create("Individual", Component3D, fitness=creator.Fitness)
toolbox = base.Toolbox()
toolbox.register("individual",
creator.Individual,
number_of_rays=number_of_rays,
ray_distribution=ray_distribution,
base_length=env.road_length,
base_width=env.road_width)
toolbox.register("population", tools.initRepeat, list, toolbox.individual)
toolbox.register("select", tools.selTournament, tournsize=2)
# Initiating first population
pop = toolbox.population(n=population_size)
# Evaluating fitness
fitnesses = []
for item in pop:
fitnesses.append(evaluate(item, env))
for ind, fit in zip(pop, fitnesses):
ind.fitness = fit
# Rendering individuals in initial population as images
for i in range(len(pop)):
draw_road(pop[i], f"{i}")
print("Drawing initial population finished")
stats_line = f"0, {max(fitnesses)}, {sum(fitnesses)/population_size}"
log_stats_init(f"stats", stats_line)
# Initiating elitism
hof = HallOfFame(1)
print("Start of evolution")
# Begin the evolution
for g in range(number_of_generations):
# A new generation
print(f"-- Generation {g} --")
# Select the next generation individuals
offspring = toolbox.select(pop, len(pop))
# Clone the selected individuals
offspring = list(map(toolbox.clone, offspring))
# Apply crossover and mutation on the offspring
for child1, child2 in zip(offspring[::2], offspring[1::2]):
# cross two individuals with probability xover_prob
if random.random() < xover_prob:
mate(child1, child2)
# fitness values of the children must be recalculated later
child1.fitness = 0
child2.fitness = 0
for mutant in offspring:
if env.configuration == "multiple free":
if random.random() < shift_segment_prob:
mutant.reflective_segments = shift_one_segment(
mutant.reflective_segments, "x")
mutant.fitness = 0
if random.random() < shift_segment_prob:
mutant.reflective_segments = shift_one_segment(
mutant.reflective_segments, "y")
mutant.fitness = 0
if random.random() < rotate_segment_prob:
mutant.reflective_segments = rotate_one_segment(
mutant.reflective_segments)
mutant.fitness = 0
if random.random() < resize_segment_prob:
mutant.reflective_segments = resize_one_segment(
mutant.reflective_segments)
mutant.fitness = 0
if env.configuration == "two connected":
if random.random() < mut_angle_prob:
mutate_angle(mutant)
mutant.fitness = 0
if random.random() < mut_length_prob:
mutate_length(mutant, length_upper_bound,
length_lower_bound)
mutant.fitness = 0
# Evaluate the individuals with an invalid fitness
invalid_ind = [ind for ind in offspring if ind.fitness == 0]
fitnesses = []
for item in invalid_ind:
fitnesses.append(evaluate(item, env))
for ind, fit in zip(invalid_ind, fitnesses):
ind.fitness = fit
print(f" Evaluated {len(invalid_ind)} individuals")
# The population is entirely replaced by the offspring
pop[:] = offspring
hof.update(pop)
# best_ind = tools.selBest(pop, 1)[0]
best_ind = hof[0]
fitnesses = []
for item in pop:
fitnesses.append(evaluate(item, env))
print(fitnesses)
stats_line = f"{g+1}, {best_ind.fitness}, {sum(fitnesses) / population_size}"
log_stats_append(f"stats", stats_line)
print(f"Best individual has fitness: {best_ind.fitness}")
draw(best_ind, f"best{g}", env)
print("-- End of (successful) evolution --")
print("--")
def create_halloffame(maxsize, rel_tol=1e-6):
return HallOfFame(maxsize, similar=EqualIndividual(rel_tol))
def evolution(
env: Environment, number_of_rays: int, ray_distribution: str,
angle_lower_bound: int, angle_upper_bound: int,
length_lower_bound: int, length_upper_bound: int,
no_of_reflective_segments: int, distance_limit: int, length_limit: int,
population_size: int, number_of_generations: int, xover_prob: float,
mut_angle_prob: float, mut_length_prob: float,
shift_segment_prob: float, rotate_segment_prob: float,
resize_segment_prob: float, tilt_base_prob: float, base_length: int,
base_slope: int, base_angle_limit_min: int, base_angle_limit_max: int):
# Initiating evolutionary algorithm
creator.create("Fitness", base.Fitness, weights=(1.0, ))
base.Fitness.weights = (1.0, 10.0, 5.0, -1.0)
creator.create("Individual", Component, fitness=creator.Fitness)
toolbox = base.Toolbox()
toolbox.register("individual",
creator.Individual,
env=env,
number_of_rays=number_of_rays,
ray_distribution=ray_distribution,
angle_lower_bound=angle_lower_bound,
angle_upper_bound=angle_upper_bound,
length_lower_bound=length_lower_bound,
length_upper_bound=length_upper_bound,
no_of_reflective_segments=no_of_reflective_segments,
distance_limit=distance_limit,
length_limit=length_limit,
base_length=base_length,
base_slope=base_slope)
toolbox.register("population", tools.initRepeat, list, toolbox.individual)
toolbox.register("evaluate", evaluate)
if env.quality_criterion == "nsgaii":
toolbox.register("select", tools.selNSGA2)
else:
toolbox.register("select", tools.selTournament, tournsize=2)
# Initiating first population
pop = toolbox.population(n=population_size)
# Evaluating fitness
fitnesses = []
for item in pop:
fitnesses.append(evaluate(item, env))
for ind, fit in zip(pop, fitnesses):
ind.fitness = fit
if env.quality_criterion != "nsgaii":
if env.configuration == "two connected":
stats_line = f"generation, best fitness, average fitness, fitness array, left segment angle, " \
f"left segment length, right segment angle, right segment length \n"
else:
stats_line = f"generation, best fitness, average fitness, fitness array, reflective segments \n"
log_stats_init(f"stats", stats_line)
# Initiating elitism
if env.quality_criterion == "nsgaii":
pop = toolbox.select(pop, len(pop))
hof = tools.ParetoFront()
hof.update(pop)
else:
hof = HallOfFame(1)
hof.update(pop)
print("Start of evolution")
# Begin the evolution
for g in range(number_of_generations):
# A new generation
print(f"-- Generation {g} --")
# Select the next generation individuals
if env.quality_criterion == "nsgaii":
offspring = tools.selTournamentDCD(pop, len(pop))
offspring = [toolbox.clone(ind) for ind in offspring]
else:
offspring = toolbox.select(pop, len(pop))
offspring = list(map(toolbox.clone, offspring))
# Apply crossover and mutation on the offspring
for child1, child2 in zip(offspring[::2], offspring[1::2]):
# cross two individuals with probability xover_prob
if env.configuration == "multiple free":
if random.random() < xover_prob:
x_over_multiple_segments(child1, child2)
# fitness values of the children must be recalculated later
child1.fitness = None
child2.fitness = None
if env.configuration == "two connected":
if random.random() < xover_prob:
x_over_two_segments(child1, child2)
# fitness values of the children must be recalculated later
child1.fitness = None
child2.fitness = None
for mutant in offspring:
if env.configuration == "multiple free":
if random.random() < shift_segment_prob:
mutant.reflective_segments = shift_one_segment(
mutant.reflective_segments, "x")
mutant.fitness = None
if random.random() < shift_segment_prob:
mutant.reflective_segments = shift_one_segment(
mutant.reflective_segments, "y")
mutant.fitness = None
if random.random() < rotate_segment_prob:
mutant.reflective_segments = rotate_one_segment(
mutant.reflective_segments)
mutant.fitness = None
if random.random() < resize_segment_prob:
mutant.reflective_segments = resize_one_segment(
mutant.reflective_segments)
mutant.fitness = None
if random.random() < tilt_base_prob:
mutant.base_slope = tilt_base(mutant.base_slope,
base_angle_limit_min,
base_angle_limit_max)
mutant.calculate_base()
mutant.original_rays = mutant.sample_rays(
number_of_rays, ray_distribution)
mutant.fitness = None
if env.configuration == "two connected":
if random.random() < mut_angle_prob:
mutate_angle(mutant)
mutant.fitness = None
if random.random() < mut_length_prob:
mutate_length(mutant, length_upper_bound,
length_lower_bound)
mutant.fitness = None
# Evaluate the individuals with an invalid fitness
invalid_ind = [ind for ind in offspring if ind.fitness is None]
fitnesses = []
for item in invalid_ind:
fitnesses.append(evaluate(item, env))
for ind, fit in zip(invalid_ind, fitnesses):
ind.fitness = fit
if env.quality_criterion == "nsgaii":
pop = toolbox.select(offspring + pop, population_size)
else:
pop[:] = offspring
hof.update(pop)
best_ind = hof[0]
fitnesses = []
for item in pop:
fitnesses.append(item.fitness)
print(fitnesses)
if env.configuration == "two connected" and env.quality_criterion != "nsgaii":
stats_line = f"{g+1}, {best_ind.fitness}, {sum(fitnesses) / population_size}, {best_ind.fitness_array}, " \
f"left angle: {180-best_ind.left_angle+best_ind.base_slope}, " \
f"left length: {best_ind.left_length_coef*best_ind.base_length}, " \
f"right angle: {best_ind.right_angle-best_ind.base_slope}, " \
f"right length: {best_ind.right_length_coef*best_ind.base_length} "
log_stats_append(f"stats", stats_line)
if env.configuration == "multiple free" and env.quality_criterion != "nsgaii":
stats_line = f"{g + 1}, {best_ind.fitness}, {sum(fitnesses) / population_size}, {best_ind.fitness_array}, "
for reflective_segment in best_ind.reflective_segments:
dimensions = f" start: {reflective_segment.p1}, end: {reflective_segment.p2}"
stats_line = stats_line + dimensions
log_stats_append(f"stats", stats_line)
print(f"Best individual has fitness: {best_ind.fitness}")
draw(best_ind, f"best{g}", env)
if env.quality_criterion == "nsgaii":
unique = choose_unique(hof, env.configuration)
stats_line = f"index, fitness array"
log_stats_init(f"stats", stats_line)
for index in range(len(unique)):
draw(unique[index], f"unique{index}", env)
if env.configuration == "two connected":
stats_line = f"{index}, {unique[index].fitness}," \
f"left angle: {180-unique[index].left_angle+unique[index].base_slope}, " \
f"left length: {unique[index].left_length_coef*unique[index].base_length}, " \
f"right angle: {unique[index].right_angle-unique[index].base_slope}, " \
f"right length: {unique[index].right_length_coef*unique[index].base_length} "
else:
stats_line = f"{index}, {unique[index].fitness}, {unique[index].base_slope}"
for reflective_segment in unique[index].reflective_segments:
dimensions = f" start: {reflective_segment.p1}, end: {reflective_segment.p2}"
stats_line = stats_line + dimensions
log_stats_append(f"stats", stats_line)
print("-- End of (successful) evolution --")
print("--")
class BaseLoop(Toolbox):
"""Base loop"""
def __init__(self, pset, pop=500, gen=20, mutate_prob=0.5, mate_prob=0.8, hall=1, re_hall=None,
re_Tree=None, initial_min=None, initial_max=3, max_value=5,
scoring=(r2_score,), score_pen=(1,), filter_warning=True, cv=1,
add_coef=True, inter_add=True, inner_add=False, vector_add=False,
cal_dim=False, dim_type=None, fuzzy=False, n_jobs=1, batch_size=40,
random_state=None, stats=None, verbose=True,
tq=True, store=False, personal_map=False, stop_condition=None):
"""
Parameters
----------
pset:SymbolSet
the feature x and traget y and others should have been added.
pop:int
number of popolation
gen:int
number of generation
mutate_prob:float
probability of mutate
mate_prob:float
probability of mate(crossover)
initial_max:int
max initial size of expression when first producing.
initial_min : None,int
max initial size of expression when first producing.
max_value:int
max size of expression
hall:int,>=1
number of HallOfFame(elite) to maintain
re_hall:None or int>=2
Notes: only vaild when hall
number of HallOfFame to add to next generation.
re_Tree: int
number of new features to add to next generation.
0 is false to add.
personal_map:bool or "auto"
"auto" is using premap and with auto refresh the premap with individual.\n
True is just using constant premap.\n
False is just use the prob of terminals.
scoring: list of Callbale, default is [sklearn.metrics.r2_score,]
See Also sklearn.metrics
score_pen: tuple of 1, -1 or float but 0.
>0 : max problem, best is positive, worse -np.inf
<0 : min problem, best is negative, worse np.inf
Notes:
if multiply score method, the scores must be turn to same dimension in preprocessing
or weight by score_pen. Because the all the selection are stand on the mean(w_i*score_i)
Examples: [r2_score] is [1],
cv=int,sklearn.model_selection._split._BaseKFold
default =1, means not cv
filter_warning:bool
filter warning or not
add_coef:bool
add coef in expression or not.
inter_add：bool
add intercept constant or not
inner_add:bool
dd inner coeffcients or not
n_jobs:int
default 1, advise 6
batch_size:int
default 40, depend of machine
random_state:int
None,int
cal_dim:bool
excape the dim calculation
dim_type:Dim or None or list of Dim
"coef": af(x)+b. a,b have dimension,f(x) is not dnan. \n
"integer": af(x)+b. f(x) is interger dimension. \n
[Dim1,Dim2]: f(x) in list. \n
Dim: f(x) ~= Dim. (see fuzzy) \n
Dim: f(x) == Dim. \n
None: f(x) == pset.y_dim
fuzzy:bool
choose the dim with same base with dim_type,such as m,m^2,m^3.
stats:dict
details of logbook to show. \n
Map:\n
values
= {"max": np.max, "mean": np.mean, "min": np.mean, "std": np.std, "sum": np.sum}
keys
= {\n
"fitness": just see fitness[0], \n
"fitness_dim_max": max problem, see fitness with demand dim,\n
"fitness_dim_min": min problem, see fitness with demand dim,\n
"dim_is_target": demand dim,\n
"coef": dim is true, coef have dim, \n
"integer": dim is integer, \n
...
}
if stats is None, default is :\n
stats = {"fitness_dim_max": ("max",), "dim_is_target": ("sum",)} for cal_dim=True
stats = {"fitness": ("max",)} for cal_dim=False
if self-definition, the key is func to get attribute of each ind./n
Examples:
def func(ind):\n
return ind.fitness[0]
stats = {func: ("mean",), "dim_is_target": ("sum",)}
verbose:bool
print verbose logbook or not
tq:bool
print progress bar or not
store:bool or path
bool or path
stop_condition:callable
stop condition on the best ind of hall, which return bool,the true means stop loop.
Examples:
def func(ind):\n
c = ind.fitness.values[0]>=0.90
return c
"""
super(BaseLoop, self).__init__()
assert initial_max <= max_value, "the initial size of expression should less than max_value limitation"
if cal_dim:
assert all(
[isinstance(i, Dim) for i in pset.dim_ter_con.values()]), \
"all import dim of pset should be Dim object."
random.seed(random_state)
self.max_value = max_value
self.pop = pop
self.gen = gen
self.mutate_prob = mutate_prob
self.mate_prob = mate_prob
self.verbose = verbose
self.cal_dim = cal_dim
self.re_hall = re_hall
self.re_Tree = re_Tree
self.store = store
self.data_all = []
self.personal_map = personal_map
self.stop_condition = stop_condition
self.cpset = CalculatePrecisionSet(pset, scoring=scoring, score_pen=score_pen,
filter_warning=filter_warning, cal_dim=cal_dim,
add_coef=add_coef, inter_add=inter_add, inner_add=inner_add,
vector_add=vector_add, cv=cv,
n_jobs=n_jobs, batch_size=batch_size, tq=tq,
fuzzy=fuzzy, dim_type=dim_type,
)
Fitness_ = newclass.create("Fitness_", Fitness, weights=score_pen)
self.PTree = newclass.create("PTrees", SymbolTree, fitness=Fitness_)
# def produce
if initial_min is None:
initial_min = 2
self.register("genGrow", genGrow, pset=self.cpset, min_=initial_min, max_=initial_max,
personal_map=self.personal_map)
self.register("genFull", genFull, pset=self.cpset, min_=initial_min, max_=initial_max,
personal_map=self.personal_map)
self.register("gen_mu", genGrow, min_=1, max_=3, personal_map=self.personal_map)
# def selection
self.register("select", selTournament, tournsize=2)
self.register("selKbestDim", selKbestDim,
dim_type=self.cpset.dim_type, fuzzy=self.cpset.fuzzy)
self.register("selBest", selBest)
self.register("mate", cxOnePoint)
# def mutate
self.register("mutate", mutUniform, expr=self.gen_mu, pset=self.cpset)
self.decorate("mate", staticLimit(key=operator.attrgetter("height"), max_value=2 * max_value))
self.decorate("mutate", staticLimit(key=operator.attrgetter("height"), max_value=2 * max_value))
if stats is None:
if cal_dim:
stats = {"fitness_dim_max": ("max",), "dim_is_target": ("sum",)}
else:
stats = {"fitness": ("max",)}
self.stats = Statis_func(stats=stats)
logbook = Logbook()
logbook.header = ['gen'] + (self.stats.fields if self.stats else [])
self.logbook = logbook
if hall is None:
hall = 1
self.hall = HallOfFame(hall)
if re_hall is None:
self.re_hall = None
else:
if re_hall is 1 or re_hall is 0:
print("re_hall should more than 1")
re_hall = 2
assert re_hall >= hall, "re_hall should more than hall"
self.re_hall = HallOfFame(re_hall)
def varAnd(self, *arg, **kwargs):
return varAnd(*arg, **kwargs)
def to_csv(self, data_all):
if self.store:
if isinstance(self.store, str):
path = self.store
else:
path = os.getcwd()
file_new_name = "_".join((str(self.pop), str(self.gen),
str(self.mutate_prob), str(self.mate_prob),
str(time.time())))
try:
st = Store(path)
st.to_csv(data_all, file_new_name)
print("store data to ", path, file_new_name)
except (IOError, PermissionError):
st = Store(os.getcwd())
st.to_csv(data_all, file_new_name)
print("store data to ", os.getcwd(), file_new_name)
def maintain_halls(self, population):
if self.re_hall is not None:
maxsize = max(self.hall.maxsize, self.re_hall.maxsize)
if self.cal_dim:
inds_dim = self.selKbestDim(population, maxsize)
else:
inds_dim = self.selBest(population, maxsize)
self.hall.update(inds_dim)
self.re_hall.update(inds_dim)
sole_inds = [i for i in self.re_hall.items if i not in inds_dim]
inds_dim.extend(sole_inds)
else:
if self.cal_dim:
inds_dim = self.selKbestDim(population, self.hall.maxsize)
else:
inds_dim = self.selBest(population, self.hall.maxsize)
self.hall.update(inds_dim)
inds_dim = []
inds_dim = copy.deepcopy(inds_dim)
return inds_dim
def re_add(self):
if self.hall.items and self.re_Tree:
it = self.hall.items
indo = it[random.choice(len(it))]
ind = copy.deepcopy(indo)
inds = ind.depart()
if not inds:
pass
else:
inds = [self.cpset.calculate_detail(indi) for indi in inds]
le = min(self.re_Tree, len(inds))
indi = inds[random.choice(le)]
self.cpset.add_tree_to_features(indi)
def re_fresh_by_name(self, *arr):
re_name = ["mutate", "genGrow", "genFull"]
if len(arr) > 0:
re_name.extend(arr)
self.refresh(re_name, pset=self.cpset)
def run(self):
# 1.generate###################################################################
population = [self.PTree(self.genFull()) for _ in range(self.pop)]
for gen_i in range(1, self.gen + 1):
population_old = copy.deepcopy(population)
# 2.evaluate###############################################################
invalid_ind_score = self.cpset.parallelize_score(population_old)
for ind, score in zip(population_old, invalid_ind_score):
ind.fitness.values = tuple(score[0])
ind.y_dim = score[1]
ind.dim_score = score[2]
population = population_old
# 3.log###################################################################
# 3.1.log-print##############################
record = self.stats.compile(population) if self.stats else {}
self.logbook.record(gen=gen_i, **record)
if self.verbose:
print(self.logbook.stream)
# 3.2.log-store##############################
if self.store:
datas = [{"gen": gen_i, "name": str(gen_i), "value": str(gen_i.fitness.values),
"dimension": str(gen_i.y_dim),
"dim_score": str(gen_i.dim_score)} for gen_i in population]
self.data_all.extend(datas)
# 3.3.log-hall###############################
inds_dim = self.maintain_halls(population)
# 4.refresh################################################################
# 4.1.re_update the premap ##################
if self.personal_map is "auto":
[self.cpset.premap.update(indi, self.cpset) for indi in inds_dim]
# 4.2.re_add_tree and refresh pset###########
if self.re_Tree:
self.re_add()
self.re_fresh_by_name()
# 5.break#######################################################
if self.stop_condition is not None:
if self.stop_condition(self.hall.items[0]):
break
# 6.next generation#######################################################
# selection and mutate,mate
population = self.select(population, len(population) - len(inds_dim))
offspring = self.varAnd(population, self, self.mate_prob, self.mutate_prob)
offspring.extend(inds_dim)
population[:] = offspring
# 7.store#####################################################################
if self.store:
self.to_csv(self.data_all)
self.hall.items = [self.cpset.calculate_detail(indi) for indi in self.hall.items]
return self.hall