def get_action_and_value_from_network(self):
if da3c_config.config.use_lstm:
action, value, lstm_state = \
self.session.op_get_action_value_and_lstm_state(state=[self.observation.queue],
lstm_state=self.lstm_state,
lstm_step=[1])
condition = self.experience is not None and (len(
self.experience) == da3c_config.config.batch_size
or self.terminal)
if not condition:
self.lstm_state = lstm_state
else:
action, value = self.session.op_get_action_and_value(
state=[self.observation.queue])
value, = value
if len(action) == 1:
if M:
self.metrics.histogram('action', action)
self.last_probs, = action
return utils.choose_action_descrete(self.last_probs), value
mu, sigma2 = action
self.last_probs = mu
if M:
self.metrics.histogram('mu', mu)
self.metrics.histogram('sigma2', sigma2)
return utils.choose_action_continuous(
mu, sigma2, da3c_config.config.output.action_low,
da3c_config.config.output.action_high), value
def action_from_policy(self, state):
assert state is not None
state = np.asarray(state)
state = np.reshape(state, (1, ) + state.shape)
probabilities, = self.session.op_get_action(state=state)
return utils.choose_action_descrete(probabilities, self.exploit)
def action_from_policy(self, state):
probabilities, = self.session.op_get_action(state=[state])
return utils.choose_action_descrete(probabilities, self.exploit)