def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.randomizer = Randomizer(Random())
selector_configuration = [
EvolutionRandomSelectorConfiguration.create(),
EvolutionRouletteSelectorConfiguration.create()
]
self.configuration = EvolutionConfiguration.create(
selector_configuration, inversion_chance=0, mutation_chance=0, crossover_chance=0)
self.create_rule_population()
self.create_grammar_statistics()
self.create_rule_adding()
self.sut = EvolutionService(self.randomizer)
self.rules = [
Rule(Symbol('S'), Symbol('NP'), Symbol('VP')),
Rule(Symbol('VP'), Symbol('VP'), Symbol('PP')),
Rule(Symbol('VP'), Symbol('V'), Symbol('NP')),
TerminalRule(Symbol('VP'), Symbol('eats')),
Rule(Symbol('PP'), Symbol('P'), Symbol('NP')),
Rule(Symbol('NP'), Symbol('Det'), Symbol('N')),
TerminalRule(Symbol('NP'), Symbol('she')), TerminalRule(Symbol('V'), Symbol('eats')),
TerminalRule(Symbol('P'), Symbol('with')), TerminalRule(Symbol('N'), Symbol('fish')),
TerminalRule(Symbol('N'), Symbol('fork')), TerminalRule(Symbol('Det'), Symbol('a'))
]
def __init__(self, randomizer, run_no, cyk_service_variant=None):
self.cyk_service_variant = cyk_service_variant if cyk_service_variant is not None \
else CykServiceVariationManager(False)
self.randomizer = randomizer
self.configuration = None
self.rule_adding = AddingRuleSupervisor.default(randomizer)
self.grammar_estimator = None
self.induction = self.cyk_service_variant.create_cyk_service(randomizer, self.rule_adding)
self.evolution = EvolutionService(randomizer)
self.stop_criteria = [NoStopCriteriaSpecified()]
self.run_no = run_no
class GcsRunner(object):
def __init__(self, randomizer, run_no, cyk_service_variant=None):
self.cyk_service_variant = cyk_service_variant if cyk_service_variant is not None \
else CykServiceVariationManager(False)
self.randomizer = randomizer
self.configuration = None
self.rule_adding = AddingRuleSupervisor.default(randomizer)
self.grammar_estimator = None
self.induction = self.cyk_service_variant.create_cyk_service(randomizer, self.rule_adding)
self.evolution = EvolutionService(randomizer)
self.stop_criteria = [NoStopCriteriaSpecified()]
self.run_no = run_no
def create_stop_criteria(self):
self.stop_criteria = [
FitnessStopCriteria(self.grammar_estimator, self.configuration),
StepStopCriteria(self.configuration),
TimeStopCriteria(self.configuration)
]
def _random_symbol_id(self, configuration):
return self.randomizer.randint(
RulePopulation.symbol_shift(),
RulePopulation.symbol_shift() + configuration.rule.max_non_terminal_symbols)
def generate_random_rules(self, provided_rules):
rules = set()
rules |= set(provided_rules)
while len(rules) < self.configuration.rule.random_starting_population_size:
rules.add(Rule(Symbol(self._random_symbol_id(self.configuration)),
Symbol(self._random_symbol_id(self.configuration)),
Symbol(self._random_symbol_id(self.configuration))))
return list(rules)
def add_initial_rules(self, initial_rules, rule_population, grammar_statistics):
for rule in initial_rules:
rule_population.add_rule(rule, self.randomizer)
grammar_statistics.on_added_new_rule(rule)
def perform_gcs(self, initial_rules, configuration, grammar_estimator,
grammar_statistics, sentences):
self.configuration = configuration
self.rule_adding.configuration = self.configuration.rule.adding
self.grammar_estimator = grammar_estimator
self.create_stop_criteria()
rule_population = self.cyk_service_variant.create_rule_population(
self.configuration.rule.starting_symbol, self.configuration.rule.universal_symbol,
max_non_terminal_symbols=self.configuration.rule.max_non_terminal_symbols)
self.add_initial_rules(self.generate_random_rules(initial_rules), rule_population,
grammar_statistics)
evolution_step = 0
# print('')
while not any(cr() for cr in self.stop_criteria):
# print('.', end='')
evolution_step_estimator = EvolutionStepEstimator()
self.induction.perform_cyk_for_all_sentences(rule_population, sentences,
evolution_step_estimator,
self.configuration.induction,
grammar_statistics)
self.grammar_estimator.append_step_estimation(evolution_step, evolution_step_estimator)
if self.configuration.should_run_evolution:
self.evolution.run_genetic_algorithm(grammar_statistics, rule_population,
self.rule_adding, self.configuration.evolution)
evolution_step += 1
self._post_step_actions(evolution_step)
stop_reasoning = next(cr for cr in self.stop_criteria if cr.has_been_fulfilled())
fitness_reached = self.grammar_estimator['fitness'].get_global_max()
# for x in rule_population.get_all_non_terminal_rules():
# print(x)
# for x in rule_population.get_terminal_rules():
# print(x)
return rule_population, stop_reasoning, fitness_reached, evolution_step
def _post_step_actions(self, step):
pass
class TestEvolution(unittest.TestCase):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.randomizer = Randomizer(Random())
selector_configuration = [
EvolutionRandomSelectorConfiguration.create(),
EvolutionRouletteSelectorConfiguration.create()
]
self.configuration = EvolutionConfiguration.create(
selector_configuration, inversion_chance=0, mutation_chance=0, crossover_chance=0)
self.create_rule_population()
self.create_grammar_statistics()
self.create_rule_adding()
self.sut = EvolutionService(self.randomizer)
self.rules = [
Rule(Symbol('S'), Symbol('NP'), Symbol('VP')),
Rule(Symbol('VP'), Symbol('VP'), Symbol('PP')),
Rule(Symbol('VP'), Symbol('V'), Symbol('NP')),
TerminalRule(Symbol('VP'), Symbol('eats')),
Rule(Symbol('PP'), Symbol('P'), Symbol('NP')),
Rule(Symbol('NP'), Symbol('Det'), Symbol('N')),
TerminalRule(Symbol('NP'), Symbol('she')), TerminalRule(Symbol('V'), Symbol('eats')),
TerminalRule(Symbol('P'), Symbol('with')), TerminalRule(Symbol('N'), Symbol('fish')),
TerminalRule(Symbol('N'), Symbol('fork')), TerminalRule(Symbol('Det'), Symbol('a'))
]
def create_rule_population(self):
self.starting_symbol = Symbol('S')
self.rule_population = RulePopulation(self.starting_symbol)
def create_grammar_statistics(self):
configuration = ClassicalStatisticsConfiguration.default()
configuration.base_fitness = 5
configuration.classical_fitness_weight = 1
configuration.fertility_weight = 1
configuration.positive_weight = 1
configuration.negative_weight = 1
self.grammar_statistics = GrammarStatistics(
configuration, self.randomizer, ClassicRuleStatistics(),
ClassicFitness())
def create_rule_adding(self):
configuration = AddingRulesConfiguration.create(
crowding_factor=2,
crowding_size=3,
elitism_size=2,
max_non_terminal_rules=19
)
adding_strategies = [SimpleAddingRuleStrategy(),
AddingRuleWithCrowdingStrategy(),
AddingRuleWithElitismStrategy()]
self.rule_adding = AddingRuleSupervisor(self.randomizer, configuration, adding_strategies)
def simulate_induction_part_work(self, rules):
for rule in rules:
self.rule_adding.add_rule(rule, self.rule_population, self.grammar_statistics)
for rule in rules:
rule_usage_info = ClassicRuleUsageInfo(True, 1)
positive_usages = self.randomizer.randint(1, 4)
for _ in range(positive_usages):
self.grammar_statistics.on_rule_usage(rule, rule_usage_info)
rule_usage_info.positive_sentence = False
self.grammar_statistics.on_rule_usage(rule, rule_usage_info)
self.grammar_statistics.fitness.get(self.grammar_statistics, rule)
def get_symbols_from_rules(self, rules):
return {y for y in chain.from_iterable(
(x.parent, x.left_child, x.right_child) for x in rules)}
def assert_contains_rules(self, rules, rules_pool):
for rule in rules:
assert_that(rules_pool, has_item(rule))
def count_rules_that_has_changed(self, old):
changed_rules = 0
for rule in self.rule_population.get_all_non_terminal_rules():
changed_rules += 1 if rule not in old else 0
return changed_rules
def test_given_no_operator_used_rule_population_should_remain_unchanged(self):
# Given:
self.simulate_induction_part_work(self.rules)
old_population = copy.deepcopy(self.rule_population)
# When:
self.sut.run_genetic_algorithm(self.grammar_statistics, self.rule_population,
self.rule_adding, self.configuration)
# Then:
self.assert_contains_rules(self.rule_population.get_all_non_terminal_rules(),
list(old_population.get_all_non_terminal_rules()))
old_symbols = self.get_symbols_from_rules(old_population.get_all_non_terminal_rules())
new_symbols = self.get_symbols_from_rules(self.rule_population.get_all_non_terminal_rules())
assert_that(old_symbols, has_items(*new_symbols))
def promote_those_rules_to_elite(self, rules):
for rule in rules:
rule_usage_info = ClassicRuleUsageInfo(True, 1)
positive_usages = 5
for _ in range(positive_usages):
self.grammar_statistics.on_rule_usage(rule, rule_usage_info)
def test_given_high_operator_usage_rule_population_should_expect_some_major_change(self):
# Given:
self.configuration.operators.inversion.chance = 1
self.configuration.operators.mutation.chance = 1
self.configuration.operators.crossover.chance = 1
self.simulate_induction_part_work(self.rules)
self.promote_those_rules_to_elite(self.rules[0:2])
elite = list(self.rules[0:2])
old_population = copy.deepcopy(self.rule_population)
# When:
self.sut.run_genetic_algorithm(self.grammar_statistics, self.rule_population,
self.rule_adding, self.configuration)
# Then:
self.assert_contains_rules(elite, list(old_population.get_all_non_terminal_rules()))
assert_that(self.count_rules_that_has_changed(old_population.get_all_non_terminal_rules()),
is_(greater_than_or_equal_to(2)))
old_symbols = self.get_symbols_from_rules(old_population.get_all_non_terminal_rules())
new_symbols = self.get_symbols_from_rules(self.rule_population.get_all_non_terminal_rules())
assert_that(old_symbols, not_(has_items(*new_symbols)))
