def cover(p0: Perception, action: int, p1: Perception, time: int,
cfg: Configuration) -> Classifier:
"""
Covering - creates a classifier that anticipates a change correctly.
The reward of the new classifier is set to 0 to prevent *reward bubbles*
in the environmental model.
Parameters
----------
p0: Perception
previous perception
action: int
chosen action
p1: Perception
current perception
time: int
current epoch
cfg: Configuration
algorithm configuration class
Returns
-------
Classifier
new classifier
"""
# In paper it's advised to set experience and reward of newly generated
# classifier to 0. However in original code these values are initialized
# with defaults 1 and 0.5 correspondingly.
new_cl = Classifier(action=action, experience=0, reward=0, cfg=cfg)
new_cl.tga = time
new_cl.talp = time
new_cl.specialize(p0, p1)
return new_cl
def test_should_return_latest_action(self, cfg):
# given
all_actions = cfg.number_of_possible_actions
population = ClassifiersList()
c0 = Classifier(action=0, cfg=cfg)
c0.talp = 1
# when
population.append(c0)
# Should return first action with no classifiers
assert 1 == choose_latest_action(population, all_actions)
# Add rest of classifiers
population.append(Classifier(action=3, cfg=cfg))
population.append(Classifier(action=7, cfg=cfg))
population.append(Classifier(action=5, cfg=cfg))
population.append(Classifier(action=1, cfg=cfg))
population.append(Classifier(action=4, cfg=cfg))
population.append(Classifier(action=2, cfg=cfg))
population.append(Classifier(action=6, cfg=cfg))
# Assign each classifier random talp from certain range
for cl in population:
cl.talp = randint(70, 100)
# But third classifier (action 7) will be the executed long time ago
population[2].talp = randint(10, 20)
# then
assert choose_latest_action(population, all_actions) == 7
