def state_update(self, obs, action):
# update history
if self.dst:
self.dst.state['history'].append([str(action)])
# NLU parsing
input_act = self.nlu.parse(
obs,
sum(self.dst.state['history'], [])
if self.dst else []) if self.nlu else obs
# state tracking
state = self.dst.update(input_act) if self.dst else input_act
# update history
if self.dst:
self.dst.state['history'][-1].append(str(obs))
# encode state
encoded_state = self.state_encoder.encode(
state) if self.state_encoder else state
if self.nlu and self.dst:
self.dst.state['user_action'] = input_act
elif self.dst and not isinstance(
self.dst, word_dst.MDBTTracker): # for act-in act-out agent
self.dst.state['user_action'] = obs
logger.nl(f'User utterance: {obs}')
logger.act(f'Inferred user action: {input_act}')
logger.state(f'Dialog state: {state}')
return input_act, state, encoded_state
def act(self, obs):
'''Standard act method from algorithm.'''
action = self.algorithm.act(self.body.encoded_state)
self.body.action = action
output_act, decoded_action = self.action_decode(action, self.body.state)
logger.act(f'System action: {action}')
logger.nl(f'System utterance: {decoded_action}')
return decoded_action
def reset(self, train_mode, config):
self.simulator.init_session()
self.history = []
user_response, user_act, session_over, reward = self.simulator.response("null", self.history)
self.last_act = user_act
logger.act(f'User action: {user_act}')
self.history.extend(["null", f'{user_response}'])
self.env_info = [State(user_response, 0., session_over)]
# update evaluator
if self.evaluator:
self.evaluator.add_goal(self.get_goal())
logger.act(f'Goal: {self.get_goal()}')
return self.env_info
def step(self, action):
user_response, user_act, session_over, reward = self.simulator.response(action, self.history)
self.last_act = user_act
self.history.extend([f'{action}', f'{user_response}'])
logger.act(f'Inferred system action: {self.get_sys_act()}')
# update evaluator
if self.evaluator:
self.evaluator.add_sys_da(self.get_sys_act())
self.evaluator.add_usr_da(self.get_last_act())
if session_over:
reward = 2.0 * self.simulator.policy.max_turn if self.evaluator.task_success() else -1.0 * self.simulator.policy.max_turn
else:
reward = -1.0
self.env_info = [State(user_response, reward, session_over)]
return self.env_info