def test_preprocess_obs(simple_obs_state_processor: SimpleObsStateProcessor,
gamestate_stub: GameState,
available_pathogens: List[Pathogen]):
expected_city_obs = []
for city in gamestate_stub.cities:
location = (np.array([90], dtype=np.float32),
np.array([-90], dtype=np.float32))
population = np.array([500], dtype=np.uint32)
connections = np.int64(2)
attributes = np.array([0, 1, 2, -2], dtype=np.int8)
pathogens = simple_obs_state_processor._build_pathogen_obs_representation(
city.pathogens, city.population,
list(reversed(available_pathogens)), gamestate_stub)
expected_city_obs.append(
(location, population, connections, attributes, pathogens))
actual_city_obs = simple_obs_state_processor.preprocess_obs(gamestate_stub)
assert compare(actual_city_obs, expected_city_obs)
class ReinforcedApproach(Approach):
trial_max = 10
def __init__(self, trainer: Trainer.__class__, weights: str):
if not ray.is_initialized():
ray.init()
self.obs_state_processor = SimpleObsStateProcessor(
infected_population_sorting_per_city)
self.act_state_processor = SimpleActStateProcessor(
sort_pathogens=self.obs_state_processor.sort_pathogens)
register_env(
"ic20env", lambda _: InferenceIC20Environment(
self.obs_state_processor, self.act_state_processor))
self.trainer = self._load_trainer(trainer(env="ic20env"), weights)
def process_round(self, state: GameState):
return self._choose_actionable_action(state)
def _choose_actionable_action(self, state: GameState) -> Action:
processed_state = self.obs_state_processor.preprocess_obs(state)
mapped_action = INVALID_ACTION
trial_count = 0
while (mapped_action == INVALID_ACTION or mapped_action.cost > state.points) \
or mapped_action not in actions.generate_possible_actions(state):
action = self.trainer.compute_action(observation=processed_state)
mapped_action, _ = self.act_state_processor.map_action(
action, state)
trial_count += 1
if trial_count >= self.trial_max:
mapped_action = actions.end_round()
break
return mapped_action
@classmethod
def _load_trainer(cls, trainer: Trainer, weights_path: str) -> Trainer:
trainer.restore(weights_path)
return trainer