def select_action(self, observation):
stpt_actions = []
therm_actions = []
blind_actions = []
idxs = []
for i in range(0, self.num_sat_actions):
rand = np.random.choice(self.stpt_action_space)
stpt_actions.append(rand)
idxs.append(np.where(self.stpt_action_space == rand)[0].item())
for i in range(0, self.num_therm_actions):
rand = np.random.choice(self.therm_action_space)
therm_actions.append(rand)
idxs.append(np.where(self.therm_action_space == rand)[0].item())
for i in range(0, self.num_blind_actions):
rand = np.random.choice(self.blind_action_space)
blind_actions.append(rand)
idxs.append(np.where(self.blind_action_space == rand)[0].item())
sat_actions_tups = []
for a in stpt_actions:
action_stpt, sat_sp = augment_ma(observation, a)
sat_actions_tups.append((action_stpt, sat_sp))
for i in range(len(idxs)):
observation[1][f"Action idx {i}"] = idxs[i]
return sat_actions_tups, therm_actions, blind_actions, idxs
def agent_start(self, state):
action, blind_action = self.policy_old.act(state[0], self.memory)
self.last_action = action
self.last_state = state
if self.discrete:
action = self.action_space[action]
blind_action = self.blind_action_space[blind_action]
action_stpt, sat_sp = augment_ma(state, action)
return action_stpt, sat_sp, blind_action
def choose_action(self, observation):
if np.random.random() > self.epsilon:
state, _, _ = observation
actions = self.q_eval.forward(state)
action_idx = T.argmax(actions).item()
action = self.action_space[action_idx]
else:
action = np.random.choice(self.action_space)
action, sat_sp = augment_ma(observation, action)
return action, sat_sp
def choose_action(self, observation):
if np.random.random() > self.epsilon:
# state = T.tensor([observation],dtype=T.float).to(self.q_eval.device)
# actions = self.q_eval.forward(state)
# action = T.argmax(actions).item()
state, _, _ = observation
actions = self.q_eval.forward(state)
action_idx = T.argmax(actions).item()
action = self.action_space[action_idx]
else:
action = np.random.choice(self.action_space)
action, sat_sp = augment_ma(observation, action)
return action, sat_sp
def select_action(self, observation, evaluate=False):
"""
:param observation:
:param evaluate:
:return: action is for the reward function
sat_sp is whats used by env_step and by the model for training
"""
actions = []
if self.start_steps > self.total_numsteps:
for i in range(self.action_space.shape[0]):
a = np.random.uniform(self.action_space[i].min(),
self.action_space[i].max())
actions.append(a)
else:
state, _, _ = observation
state = torch.FloatTensor(state.float()).to(
self.device).unsqueeze(0)
if evaluate is False:
action, _, _ = self.policy.sample(state)
else:
_, _, action = self.policy.sample(state)
for i in range(self.action_space.shape[0]):
a = action.detach().cpu().numpy()[0][i]
actions.append(a)
sat_actions = actions[0:self.num_sat_actions]
therm_actions = actions[self.num_sat_actions:self.num_therm_actions +
self.num_sat_actions]
blind_actions = actions[self.num_therm_actions + self.num_sat_actions:]
sat_actions_tups = []
for a in sat_actions:
action_stpt, sat_sp = augment_ma(observation, a)
sat_actions_tups.append((action_stpt, sat_sp))
if len(sat_actions) == 0:
# this is hacky but makes the parsing in the main file cleaner
sat_actions_tups.append(([], []))
return sat_actions_tups, therm_actions, blind_actions, actions
def select_action(self, observation):
if np.random.random() > self.epsilon:
state, _, _ = observation
actions_stpt, actions_blinds = self.q_eval.forward(state)
# setpoint actions
actions_stpt_idx = T.argmax(actions_stpt).item()
action_stpt = self.action_space[0][actions_stpt_idx]
# blind actions
actions_blinds_idx = T.argmax(actions_blinds).item()
action_blinds = self.action_space[1][actions_blinds_idx]
else:
action_stpt = np.random.choice(self.action_space[0])
action_blinds = np.random.choice(self.action_space[1])
action_stpt, sat_sp = augment_ma(observation, action_stpt)
return action_stpt, sat_sp, action_blinds
def agent_start(self, state):
action_idx = self.policy_old.act(state[0], self.memory)
self.last_action = action_idx[:]
self.last_state = state
actions = list()
for i, action in enumerate(action_idx):
actions.append(self.action_space[i][action])
sat_actions = actions[:self.num_sat_actions]
therm_actions = actions[self.num_sat_actions:self.num_therm_actions +
self.num_sat_actions]
blind_actions = actions[self.num_therm_actions + self.num_sat_actions:]
sat_actions_tups = []
for action in sat_actions:
action_stpt, sat_sp = augment_ma(state, action)
sat_actions_tups.append((action_stpt, sat_sp))
if len(sat_actions) == 0:
# this is hacky but makes the parsing in the main file cleaner
sat_actions_tups.append(([], []))
return sat_actions_tups, therm_actions, blind_actions