def _run_trial_explore(self, env, trials, current_trial) -> TrialMetrics:
logger.debug("** Running trial explore ** ")
# Initial conditions
steps = 0
raw_state = env.reset()
state = self.cfg.environment_adapter.to_genotype(raw_state)
action = env.action_space.sample()
last_reward = 0
prev_state = Perception.empty()
selected_cl = None
prev_selected_cl = None
done = False
while not done:
state = Perception(state)
match_set = self.population.form_match_set(state)
if steps > 0:
alp.apply(prev_state,
state,
selected_cl,
self.population)
rl.bucket_brigade_update(
selected_cl,
prev_selected_cl,
last_reward)
prev_selected_cl = selected_cl
# TODO: you can do it better
if random.random() < self.cfg.epsilon:
selected_cl = random.choice(match_set)
else:
selected_cl = self._best_cl(match_set)
action = selected_cl.action
iaction = self.cfg.environment_adapter.to_lcs_action(action)
logger.debug("\tExecuting action: [%d]", action)
prev_state = Perception(state)
raw_state, last_reward, done, _ = env.step(iaction)
state = self.cfg.environment_adapter.to_genotype(raw_state)
state = Perception(state)
if done:
alp.apply(prev_state,
state,
selected_cl,
self.population)
rl.bucket_brigade_update(
selected_cl,
prev_selected_cl,
last_reward)
steps += 1
return TrialMetrics(steps, last_reward)
def test_should_perform_bucket_brigade_update_when_first_step(self):
# given
prev_cl = None
cl = Classifier(0.5, None)
# when
rl.bucket_brigade_update(cl, prev_cl, 100)
# then
assert cl.r == 0.5
assert prev_cl is None
def test_should_perform_bucket_brigade_update(self, _r0, reward, _r1):
# given
prev_cl = Classifier(_r0, None)
cl = Classifier(0.5, None)
# when
rl.bucket_brigade_update(cl, prev_cl, reward)
# then
assert cl.r == 0.5
assert prev_cl.r == _r1
assert cl.ir is None
assert prev_cl.ir is None