def inst_regret(self, t, all_paths, selected_path, robot_model, param = None):
''' The instantaneous Kapoor regret of a selected path, according to the specified reward function
Input:
all_paths: the set of all avalaible paths to the robot at time t
selected path: the path selected by the robot at time t
robot_model (GP Model)
'''
value_omni = {}
for path, points in all_paths.items():
if param is None:
value_omni[path] = self.f_rew(time = t, xvals = points, robot_model = robot_model)
else:
value_omni[path] = aqlib.mves(time = t, xvals = points, robot_model = robot_model, param = (self.max_val).reshape(1,1))
value_max = value_omni[max(value_omni, key = value_omni.get)]
if param is None:
value_selected = self.f_rew(time = t, xvals = selected_path, robot_model = robot_model)
else:
value_selected = aqlib.mves(time = t, xvals = selected_path, robot_model = robot_model, param = (self.max_val).reshape(1,1))
return value_max - value_selected, value_selected, value_max
def update_metrics(self, t, robot_model, all_paths, selected_path):
''' Function to update avaliable metrics'''
# Compute aquisition function
if (self.f_aqu == aqlib.mves):
self.metrics['aquisition_function'][t] = self.f_aqu(
t, selected_path, robot_model, [None])
else:
self.metrics['aquisition_function'][t] = self.f_aqu(
t, selected_path, robot_model)
# Compute reward functions
self.metrics['mean_reward'][t] = self.mean_reward(
t, selected_path, robot_model)
self.metrics['info_gain_reward'][t] = self.info_gain_reward(
t, selected_path, robot_model)
self.metrics['hotspot_info_reward'][t] = self.hotspot_info_reward(
t, selected_path, robot_model)
self.metrics['mes_reward_robot'][t] = aqlib.mves(
t, selected_path, robot_model, [None])
# Compute other performance metrics
self.metrics['MSE'][t] = self.MSE(robot_model, NTEST=25)
def playback(playback_locs, playback_samples, max_val, column_names):
''' Gather noisy samples of the environment and updates the robot's GP model
Input:
T (int > 0): the length of the planning horization (number of planning iterations)'''
d = playback_locs
data = pd.read_table(d, delimiter=" ", header=None)
data = data.T
if data.shape[1] > len(column_names):
data = pd.read_table(d, delimiter=" ", header=None, skipfooter=2)
data = data.T
data.columns = column_names
robot_loc = np.vstack((data['robot_loc_x'], data['robot_loc_y'])).T
d = playback_samples
data = pd.read_table(d, delimiter=" ", header=None)
data = data.T
data.columns = ['x1', 'x2', 'z']
sample_loc = np.vstack((data['x1'], data['x2'])).T
sample_val = data['z'].T
# Initialize the robot's GP model with the initial kernel parameters
extent = (0., 10., 0., 10.)
init_variance = 100.0
init_lengthscale = 1.0
noise = 1.001
GP = gplib.OnlineGPModel(ranges=extent,
lengthscale=init_lengthscale,
variance=init_variance,
noise=noise)
t_sample_locs = {}
t_sample_vals = {}
value_robot = []
S = 0
E = 0
for t, end_loc in enumerate(robot_loc[1:, :]):
# Get next stop point in stream
while (not np.isclose(sample_loc[E, 0], end_loc[0])
or not np.isclose(sample_loc[E, 1], end_loc[1])):
E += 1
E += 1
t_sample_locs[t] = sample_loc[S:E, :]
t_sample_vals[t] = np.array((sample_val[S:E])).astype('float')
S = E
E += 1
#print "--------------", t, "-----------------"
#print t_sample_locs[t]
#print t_sample_vals[t]
value_robot.append(
aqlib.mves(time=t,
xvals=t_sample_locs[t],
robot_model=GP,
param=(np.array(max_val)).reshape(1, 1)))
GP.add_data(
t_sample_locs[t],
np.reshape(t_sample_vals[t], (t_sample_locs[t].shape[0], 1)))
t = 149
t_sample_locs[t] = sample_loc[S:, :]
t_sample_vals[t] = np.array((sample_val[S:])).astype('float')
#print "--------------", t, "-----------------"
#print t_sample_locs[t]
#print t_sample_vals[t]
value_robot.append(
aqlib.mves(time=t,
xvals=t_sample_locs[t],
robot_model=GP,
param=(np.array(max_val)).reshape(1, 1)))
#GP.add_data(t_sample_locs[t], np.reshape(t_sample_vals[t], (t_sample_locs[t].shape[0], 1)))
return np.cumsum(value_robot)
