class DialogEnv(gym.Env):
def __init__(
self,
user_goals: List[UserGoal],
emc_params: Dict,
max_round_num: int,
database: Dict,
slot2values: Dict[str, List[Any]],
) -> None:
self.user = UserSimulator(user_goals, max_round_num)
self.emc = ErrorModelController(slot2values, emc_params)
self.state_tracker = StateTracker(database, max_round_num)
self.action_space = gym.spaces.Discrete(len(AGENT_ACTIONS))
self.observation_space = gym.spaces.multi_binary.MultiBinary(
self.state_tracker.get_state_size())
def step(self, agent_action_index: int):
agent_action = map_index_to_action(agent_action_index)
self.state_tracker.update_state_agent(agent_action)
user_action, reward, done, success = self.user.step(agent_action)
if not done:
self.emc.infuse_error(user_action)
self.state_tracker.update_state_user(user_action)
next_state = self.state_tracker.get_state(done)
return next_state, reward, done, success
def reset(self):
self.state_tracker.reset()
init_user_action = self.user.reset()
self.emc.infuse_error(init_user_action)
self.state_tracker.update_state_user(init_user_action)
return self.state_tracker.get_state()
def __init__(
self,
user_goals: List[UserGoal],
emc_params: Dict,
max_round_num: int,
database: Dict,
slot2values: Dict[str, List[Any]],
) -> None:
self.user = UserSimulator(user_goals, max_round_num)
self.emc = ErrorModelController(slot2values, emc_params)
self.state_tracker = StateTracker(database, max_round_num)
self.action_space = gym.spaces.Discrete(len(AGENT_ACTIONS))
self.observation_space = gym.spaces.multi_binary.MultiBinary(
self.state_tracker.get_state_size())
# Clean DB
remove_empty_slots(database)
# Load movie dict
db_dict = pickle.load(open(DICT_FILE_PATH, 'rb'), encoding='latin1')
# Load goal File
user_goals = pickle.load(open(USER_GOALS_FILE_PATH, 'rb'),
encoding='latin1')
# Init. Objects
if USE_USERSIM:
user = UserSimulator(user_goals, constants, database)
else:
user = User(constants)
emc = ErrorModelController(db_dict, constants)
state_tracker = StateTracker(database, constants)
sarsa_agent = SARSAgent(state_tracker.get_state_size(), constants)
#dqn_agent = DQNAgent(state_tracker.get_state_size(), constants)
def run_round(state, warmup=False):
# 1) Agent takes action given state tracker's representation of dialogue (state)
agent_action_index, agent_action = sarsa_agent.get_action(state,
use_rule=warmup)
# 2) Update state tracker with the agent's action
state_tracker.update_state_agent(agent_action)
# 3) User takes action given agent action
user_action, reward, done, success = user.step(agent_action)
if not done:
# 4) Infuse error into semantic frame level of user action
