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()
# 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
emc.infuse_error(user_action)
# 5) Update state tracker with user action
# 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)
dqn_agent = DuellingQNetworkAgent(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 = dqn_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
emc.infuse_error(user_action)
# 5) Update state tracker with user action
# Load movie dict
# db_dict = pickle.load(open(DICT_FILE_PATH, 'rb'), encoding='latin1')
db_dict = json.load(open(DICT_FILE_PATH, encoding='utf-8'))[0]
# Load goal file
# user_goals = pickle.load(open(USER_GOALS_FILE_PATH, 'rb'), encoding='latin1')
user_goals = json.load(open(USER_GOALS_FILE_PATH, encoding='utf-8'))
# 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)
dqn_agent = DQNAgent(state_tracker.get_state_size(), constants)
def test_run():
"""
Runs the loop that tests the agent.
Tests the agent on the goal-oriented chatbot task. Only for evaluating a trained agent. Terminates when the episode
reaches NUM_EP_TEST.
"""
print('Testing Started...')
episode = 0
while episode < NUM_EP_TEST:
episode_reset()
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)
sess = K.get_session()
ac_agent = ActorCritic(state_tracker.get_state_size(), constants, sess)
#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 = ac_agent.act(state)
# 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
emc.infuse_error(user_action)
# 5) Update state tracker with user action
state_tracker.update_state_user(user_action)
# 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)
acm = AdvantageACM(state_tracker.get_state_size(), constants)
def run_round(state, Agent_Actions, User_Actions, tot_slt_len, stp, q, warmup):
u_r = 0 ##User Repeatition
a_r = 0 ##Agent Repeatition
a_q = 0 ##Agent Question
u_q = q ##User Question
pen = 0 ##User asked Question Agent replied Question
# 1) Agent takes action given state tracker's representation of dialogue (state)
agent_action_index, agent_action = acm.act(state, warmup)
print('Agent Action_Index:', agent_action_index)
#print('Agent_Action:',agent_action)
if (agent_action['intent'] == 'request'):
a_q = 1
if (agent_action_index in Agent_Actions):
# 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)
dqn_agent = DRQNAgent(
state_tracker.get_state_size(),
constants) # the variable dqn agent is intialized to a DRQN agent
def run_round(states, warmup=False):
# 1) Agent takes action given state tracker's representation of dialogue (state)
#print(states[0].shape)
if len(states) > 3:
state_1 = np.stack((states[-4], states[-3], states[-2], states[-1]),
axis=0)
else:
state_1 = np.vstack((np.zeros(
(4 - len(states), 224)), np.array(states)))
agent_action_index, agent_action = dqn_agent.get_action(state_1,
use_rule=warmup)
# 2) Update state tracker with the agent's action