def test_lexicon_sample_storage_adequacy(self):
game = DiscriminationGame()
the_agent = SteelsClassificationAgent()
iris = datasets.load_iris()
the_environment = environment.Environment(iris.data, iris.target)
the_agent.add_sample(1, the_environment)
self.lexicon_and_sample_storage_adequacy(the_agent)
the_agent.forget()
self.lexicon_and_sample_storage_adequacy(the_agent)
category = the_agent.predict(the_environment.get_sample(1))
self.lexicon_and_sample_storage_adequacy(the_agent)
word = the_agent.find_word_for_category(category)
self.lexicon_and_sample_storage_adequacy(the_agent)
the_agent.forget()
self.lexicon_and_sample_storage_adequacy(the_agent)
game.learning_after_game(the_agent, 2, the_environment, category, True)
self.lexicon_and_sample_storage_adequacy(the_agent)
for _ in range(1000):
the_agent.forget()
self.lexicon_and_sample_storage_adequacy(the_agent)
assert_equal(0, the_agent.sample_storage.get_categories_size())
assert_equal(0, the_agent.lexicon.get_categories_size())
class TestDiscriminationGame:
"""
Functions tested in TestDiscriminationGame:
- interact
Functions not tested:
- __init__
- learning_after_game
- play
- sample_from_other_class
"""
def __init__(self):
self.agents = None
digits = datasets.load_digits()
self.environment = Environment(digits.data, digits.target)
self.game = DiscriminationGame()
def setup(self):
agents = []
for _ in range(10):
agent = SteelsAgent()
agent.set_fitness("DG", DummyFitness())
agents.append(agent)
self.agents = DummyNetwork(agents)
def test_interact_long_run(self):
for _ in range(100):
self.game.interact(self.agents, self.environment)
def __init__(self):
self.agents = None
digits = datasets.load_digits()
self.environment = Environment(digits.data, digits.target)
self.game = DiscriminationGame()
class GuessingGame:
"""
This class implements steels guessing game.
:param long samples_number: the number of samples shown to agent in one interaction including topic. \
Value from 2 to infinity.
:param float good_agent_measure: the threshold which affects generation of new categories. Value from 0 to 1.
:param string role_model: the way of assign agents to roles.\
"RANDOM" - both agents has equal chance of becoming speaker. \
"SPEAKER" - first agent always becomes speaker. \
"HEARER" - first agent always becomes hearer.
Guessing game was described in Luc Steels and Tony Belpaeme
"Coordinating perceptually grounded categories through language: A case study for colour".
"""
def __init__(self, samples_number=4, good_agent_measure=0.95, role_model="RANDOM"):
self.samples_number = samples_number
self.good_agent_measure = good_agent_measure
self.role_model = role_model
self.game = DiscriminationGame(samples_number, good_agent_measure)
def interact(self, agents, environment):
"""
One turn of interaction in guessing game.
:param Network agents: source of agents for discrimination game.
:param Environment environment: source of stimuli for discrimination game.
"""
agents = agents.get_agents(2)
# Assigning roles to agents.
if self.role_model == "RANDOM":
if random.randint(0, 1):
speaker = agents[0]
hearer = agents[1]
else:
speaker = agents[1]
hearer = agents[0]
elif self.role_model == "SPEAKER":
speaker = agents[0]
hearer = agents[1]
elif self.role_model == "HEARER":
speaker = agents[1]
hearer = agents[0]
else:
raise ValueError
# Choose of topic and other stimuli from environment.
topic_index, topic, topic_class = environment.get_random_sample()
other_samples = [
DiscriminationGame.sample_from_other_class(topic_class, environment) for _ in range(self.samples_number - 1)
]
# Speaker is trying to discriminate topic from other samples.
result, speaker_category = self.game.play_with_given_samples(speaker, topic, other_samples)
# If speaker can discriminate topic from other_samples.
if result:
# Speaker is looking for word that suits topic category.
word = speaker.find_word_for_category(speaker_category)
# Hearer is looking for category that suits speaker word.
hearer_category = hearer.find_category_for_word(word)
# If hearer finds such category.
if hearer_category is not None:
# The other samples and topic are shuffled to avoid position clue.
all_samples = other_samples + [topic]
random.shuffle(all_samples)
# Hearer trying to guess topic that belongs the most to category.
guessed_topic = hearer.the_best_sample_for_category(hearer_category, all_samples)
# If hearer correctly assumed that topic belongs to guesses category the most.
if np.array_equal(topic, guessed_topic):
result = "Success"
# If hearer didn't assume that topic belongs to guesses category the most.
else:
result = "Not matching samples."
# If hearer doesn't find such category for word.
else:
result = "Hearer didn't know word."
# If speaker cannot discriminate topic from other_samples.
else:
result = "Speaker failed discrimination."
# Actions taken up depending on game results.
if result == "Success":
# Strengthen association between word and category.
speaker.increase_weight_word_category(word, speaker_category)
# Strengthen association between word and category.
hearer.increase_weight_word_category(word, hearer_category)
# Weaken associations between other categories and word.
speaker.decrease_weights_for_other_categories(word, speaker_category)
# Weaken associations between other words and category.
hearer.decrease_weights_for_other_words(word, hearer_category)
# Strengthen memory of category.
speaker.increase_weights_sample_category(topic_index, environment, speaker_category)
# Strengthen memory of category.
hearer.increase_weights_sample_category(topic_index, environment, hearer_category)
speaker.update_fitness_measure("GG", True)
speaker.update_fitness_measure("DG", True)
hearer.update_fitness_measure("GG", True)
speaker.forget()
hearer.forget()
elif result == "Speaker failed discrimination.":
# Speaker learns topic.
self.game.learning_after_game(speaker, topic_index, environment, speaker_category, False)
speaker.update_fitness_measure("GG", False)
speaker.update_fitness_measure("DG", False)
hearer.update_fitness_measure("GG", False)
speaker.forget()
elif result == "Hearer didn't know word.":
# Hearer plays discrimination game.
result, hearer_category = self.game.play_with_given_samples(
hearer, environment.get_sample(topic_index), other_samples
)
# If hearer cannot discriminate topic.
if not result:
# Hearer learns topic.
hearer_category, _ = self.game.learning_after_game(
hearer, topic_index, environment, hearer_category, result
)
# If hearer cannot learn then crash.
if hearer_category is None:
raise ValueError
# Associating the category for topic with word.
hearer.add_word_to_category(word, hearer_category)
speaker.update_fitness_measure("GG", False)
speaker.update_fitness_measure("DG", True)
hearer.update_fitness_measure("GG", False)
hearer.update_fitness_measure("DG", result)
speaker.forget()
hearer.forget()
elif result == "Not matching samples.":
# Weaken association between word and category.
speaker.decrease_weight_word_category(word, speaker_category)
# Weaken association between word and category.
hearer.decrease_weight_word_category(word, hearer_category)
# Hearer learns topic.
self.game.learning_after_game(hearer, topic_index, environment, hearer_category, result=False)
speaker.update_fitness_measure("GG", False)
speaker.update_fitness_measure("DG", True)
hearer.update_fitness_measure("GG", False)
speaker.forget()
hearer.forget()
else:
raise ValueError
def interact(self, agents, environment):
"""
One turn of interaction in guessing game.
:param Network agents: source of agents for discrimination game.
:param Environment environment: source of stimuli for discrimination game.
"""
agents = agents.get_agents(2)
# Assigning roles to agents.
if self.role_model == "RANDOM":
if random.randint(0, 1):
speaker = agents[0]
hearer = agents[1]
else:
speaker = agents[1]
hearer = agents[0]
elif self.role_model == "SPEAKER":
speaker = agents[0]
hearer = agents[1]
elif self.role_model == "HEARER":
speaker = agents[1]
hearer = agents[0]
else:
raise ValueError
# Choose of topic and other stimuli from environment.
topic_index, topic, topic_class = environment.get_random_sample()
other_samples = [
DiscriminationGame.sample_from_other_class(topic_class, environment) for _ in range(self.samples_number - 1)
]
# Speaker is trying to discriminate topic from other samples.
result, speaker_category = self.game.play_with_given_samples(speaker, topic, other_samples)
# If speaker can discriminate topic from other_samples.
if result:
# Speaker is looking for word that suits topic category.
word = speaker.find_word_for_category(speaker_category)
# Hearer is looking for category that suits speaker word.
hearer_category = hearer.find_category_for_word(word)
# If hearer finds such category.
if hearer_category is not None:
# The other samples and topic are shuffled to avoid position clue.
all_samples = other_samples + [topic]
random.shuffle(all_samples)
# Hearer trying to guess topic that belongs the most to category.
guessed_topic = hearer.the_best_sample_for_category(hearer_category, all_samples)
# If hearer correctly assumed that topic belongs to guesses category the most.
if np.array_equal(topic, guessed_topic):
result = "Success"
# If hearer didn't assume that topic belongs to guesses category the most.
else:
result = "Not matching samples."
# If hearer doesn't find such category for word.
else:
result = "Hearer didn't know word."
# If speaker cannot discriminate topic from other_samples.
else:
result = "Speaker failed discrimination."
# Actions taken up depending on game results.
if result == "Success":
# Strengthen association between word and category.
speaker.increase_weight_word_category(word, speaker_category)
# Strengthen association between word and category.
hearer.increase_weight_word_category(word, hearer_category)
# Weaken associations between other categories and word.
speaker.decrease_weights_for_other_categories(word, speaker_category)
# Weaken associations between other words and category.
hearer.decrease_weights_for_other_words(word, hearer_category)
# Strengthen memory of category.
speaker.increase_weights_sample_category(topic_index, environment, speaker_category)
# Strengthen memory of category.
hearer.increase_weights_sample_category(topic_index, environment, hearer_category)
speaker.update_fitness_measure("GG", True)
speaker.update_fitness_measure("DG", True)
hearer.update_fitness_measure("GG", True)
speaker.forget()
hearer.forget()
elif result == "Speaker failed discrimination.":
# Speaker learns topic.
self.game.learning_after_game(speaker, topic_index, environment, speaker_category, False)
speaker.update_fitness_measure("GG", False)
speaker.update_fitness_measure("DG", False)
hearer.update_fitness_measure("GG", False)
speaker.forget()
elif result == "Hearer didn't know word.":
# Hearer plays discrimination game.
result, hearer_category = self.game.play_with_given_samples(
hearer, environment.get_sample(topic_index), other_samples
)
# If hearer cannot discriminate topic.
if not result:
# Hearer learns topic.
hearer_category, _ = self.game.learning_after_game(
hearer, topic_index, environment, hearer_category, result
)
# If hearer cannot learn then crash.
if hearer_category is None:
raise ValueError
# Associating the category for topic with word.
hearer.add_word_to_category(word, hearer_category)
speaker.update_fitness_measure("GG", False)
speaker.update_fitness_measure("DG", True)
hearer.update_fitness_measure("GG", False)
hearer.update_fitness_measure("DG", result)
speaker.forget()
hearer.forget()
elif result == "Not matching samples.":
# Weaken association between word and category.
speaker.decrease_weight_word_category(word, speaker_category)
# Weaken association between word and category.
hearer.decrease_weight_word_category(word, hearer_category)
# Hearer learns topic.
self.game.learning_after_game(hearer, topic_index, environment, hearer_category, result=False)
speaker.update_fitness_measure("GG", False)
speaker.update_fitness_measure("DG", True)
hearer.update_fitness_measure("GG", False)
speaker.forget()
hearer.forget()
else:
raise ValueError
def __init__(self, samples_number=4, good_agent_measure=0.95, role_model="RANDOM"):
self.samples_number = samples_number
self.good_agent_measure = good_agent_measure
self.role_model = role_model
self.game = DiscriminationGame(samples_number, good_agent_measure)
