class Cluster(object):
def __init__(self):
self.init_states()
self.setup_network()
self.setup_parameters()
def init_states(self):
self.state_G = nx.Graph()
self.state_xcor = []
self.state_ycor = []
self.state_link = []
self.state_server = []
self.state_cluster_id = []
self.state_cluster_core_xcor = []
self.state_cluster_core_ycor = []
self.state_event_xcor = []
self.state_event_ycor = []
self.state_event_step_rate_xcor = 0
self.state_event_step_rate_ycor = 0
def reuse_network(self, s_xcor, s_ycor):
self.reuse_network_topology(s_xcor, s_ycor)
def setup_network(self):
self.set_network_topology()
while self.all_nodes_connected() == False:
self.set_network_topology()
def setup_parameters(self):
self.set_server_and_header()
self.set_cluster_id()
self.set_events()
def draw_network(self):
pos = nx.kamada_kawai_layout(self.state_G)
nx.draw(self.state_G,
pos,
with_labels=True,
cmap=plt.get_cmap('Accent'),
node_color=self.state_cluster_id,
node_size=200)
plt.show()
# Make sure the distance between neighbor nodes is larger than a required value.
def check_neighbor_distance_larger_than_min_range(self, node_id):
good_position = 1
for j in range(0, node_id):
ax = self.state_xcor[node_id]
ay = self.state_ycor[node_id]
bx = self.state_xcor[j]
by = self.state_ycor[j]
distance_square = (ax - bx)**2 + (ay - by)**2
if distance_square < min_distance_between_nodes_square:
good_position = 0
return good_position
# Deploy the nodes in random positions.
def scatter_node_random_position(self):
for i in range(0, total_node_number):
good_position = 0
for find_good_position_time in range(0,
max_find_good_position_time):
if good_position == 0:
self.state_xcor[i] = random.random() * deploy_range_x
self.state_ycor[i] = random.random() * deploy_range_y
good_position = self.check_neighbor_distance_larger_than_min_range(
i)
# The state_link is the connectivity matrix of the network.
def set_network_connectivity(self):
self.state_link = []
transmit_range_square = transmit_range**2
for i in range(0, node_number + server_number):
node_link = []
for j in range(0, node_number + server_number):
if i != j and (self.state_xcor[i] - self.state_xcor[j])**2 + (
self.state_ycor[i] -
self.state_ycor[j])**2 <= transmit_range_square:
node_link.append(1)
else:
node_link.append(0)
self.state_link.append(node_link)
self.set_graph()
def reuse_network_topology(self, s_xcor, s_ycor):
self.state_xcor = s_xcor
self.state_ycor = s_ycor
self.set_network_connectivity()
# Use the saved coordinate values or not.
def set_network_topology(self):
'''
self.state_xcor = []
self.state_ycor = []
for i in range(0, node_number+server_number):
self.state_xcor.append(0)
self.state_ycor.append(0)
'''
self.state_xcor = np.zeros((node_number + server_number, ))
self.state_ycor = np.zeros((node_number + server_number, ))
if flag_benchmark_topology == 1:
self.scatter_node_random_position()
with open('state_xcor.txt', 'w') as f:
for item in self.state_xcor:
f.write("%s\n" % item)
f.close()
with open('state_ycor.txt', 'w') as f:
for item in self.state_ycor:
f.write("%s\n" % item)
f.close()
else:
f = open("state_xcor.txt", "r")
i = 0
for x in f:
self.state_xcor[i] = float(x.strip())
i = i + 1
f.close()
f = open("state_ycor.txt", "r")
i = 0
for y in f:
self.state_ycor[i] = float(y.strip())
i = i + 1
f.close()
self.set_network_connectivity()
# The positions and ID of cluster headers are initialized.
def set_server_and_header(self):
self.state_server = []
for i in range(0, server_number):
self.state_server.append(i)
self.state_cluster_core_xcor.append(0)
self.state_cluster_core_ycor.append(0)
self.state_cluster_core_xcor[i] = self.state_xcor[i]
self.state_cluster_core_ycor[i] = self.state_ycor[i]
# Select the node that is closest to the moving cluster-core as the cluster header. Use this cluster header to make voronoi clusters.
def set_cluster_id(self):
self.state_cluster_id = [0] * (node_number + server_number)
temp_cluster_proxy = []
for i in range(0, server_number):
self.state_cluster_id[i] = self.state_server[i]
temp_cluster_proxy.append(0)
min_distance_square = (deploy_range_x + deploy_range_y)**2
min_id = -1
for j in range(server_number, node_number + server_number):
temp_distance_square = (self.state_cluster_core_xcor[i] -
self.state_xcor[j])**2 + (
self.state_cluster_core_ycor[i] -
self.state_ycor[j])**2
if temp_distance_square < min_distance_square:
min_distance_square = temp_distance_square
min_id = j
temp_cluster_proxy[i] = min_id
for i in range(server_number, node_number + server_number):
closest_header_cluster_id = -1
closest_distance = node_number
for j in range(0, server_number):
header_id = temp_cluster_proxy[j]
hop_distance = len(self.find_route(i, header_id)) - 1
if hop_distance < closest_distance:
closest_header_cluster_id = self.state_cluster_id[j]
closest_distance = hop_distance #?? here the hop_distance is not used after this
self.state_cluster_id[i] = closest_header_cluster_id
# In the network init stage, make sure all nodes are connected to the network.
def all_nodes_connected(self):
for i in range(0, node_number + server_number):
for j in range(0, node_number + server_number):
check = nx.has_path(self.state_G, i, j)
if check == False:
return False
return True
# Init graph parameters for python graph library.
def set_graph(self):
self.state_G = nx.Graph()
for i in range(0, node_number + server_number):
self.state_G.add_node(i)
for i in range(0, node_number + server_number):
for j in range(i, node_number + server_number):
if self.state_link[i][j] == 1 and self.state_link[j][i] == 1:
self.state_G.add_edge(i, j)
# Find the path from source node "s" to the destination node "t".
def find_route(self, s, t):
check = nx.has_path(self.state_G, source=s, target=t)
if check == True:
path = nx.dijkstra_path(self.state_G, source=s, target=t)
else:
path = []
return path
# Every time tick, we run this function. This function generates data from each node to the cluster servers.
# If the senser is in the detection range of moving event, the senser generates more data.
def transmit_flow_in_network(self):
state_cluster_path_length = [0] * server_number
state_cluster_size = [0] * server_number
for i in range(server_number, total_node_number):
j = self.state_cluster_id[i]
pass_route = self.find_route(i, j)
event_distance_square = (self.state_xcor[i] - self.state_event_xcor
)**2 + (self.state_ycor[i] -
self.state_event_ycor)**2
if event_distance_square < event_detection_range_square:
state_cluster_size[
j] = state_cluster_size[j] + 1 * event_data_increase_rate
state_cluster_path_length[j] = state_cluster_path_length[
j] + len(pass_route) * event_data_increase_rate
else:
state_cluster_size[j] = state_cluster_size[j] + 1
state_cluster_path_length[
j] = state_cluster_path_length[j] + len(pass_route)
# To calculate the reward values, please refer the paper.
mean_state_balance_data = np.mean(state_cluster_size)
total_state_balance = 0
for k in range(0, len(state_cluster_size)):
total_state_balance = total_state_balance + (
1 - abs(state_cluster_size[k] - mean_state_balance_data) /
mean_state_balance_data)
state_balance_mse_data = total_state_balance / len(state_cluster_size)
# To calculate the reward values, please refer the paper.
mean_state_balance_com = np.mean(state_cluster_path_length)
total_state_balance = 0
for k in range(0, len(state_cluster_path_length)):
total_state_balance = total_state_balance + (
1 - abs(state_cluster_path_length[k] - mean_state_balance_com)
/ mean_state_balance_com)
state_balance_mse_com = total_state_balance / len(
state_cluster_path_length)
return (state_balance_mse_data * state_balance_mse_com)
# This function generate a vector, each value of the vector represents how much data size the senser is producing at this time tick.
def workload_in_network(self):
state_workload = [0] * total_node_number
for i in range(server_number, total_node_number):
event_distance_square = (self.state_xcor[i] - self.state_event_xcor
)**2 + (self.state_ycor[i] -
self.state_event_ycor)**2
if event_distance_square < event_detection_range_square:
state_workload[i] = 1 * event_data_increase_rate
else:
state_workload[i] = 1
return state_workload
# If in the gameover mode, we check how much the clusters are unbalanced. If the unbalanced value is over a threshold, then game over.
def check_flow_in_network_fail(self):
balance = self.transmit_flow_in_network()
if balance < max_balance_diff:
return 1
else:
return 0
# Make sure the graph matrix is correctly created. The network is bi-direction, so the matrix should be symmetrical。
def check_graph_error(self):
for i in range(0, node_number + server_number):
for j in range(i, node_number + server_number):
if self.state_link[i][j] != self.state_link[j][i]:
print("Error: node network topology.")
exit()
# For each action number, move the position of cluster-core.
def action_route(self, action, tick):
# [5 actions]: stay, up, down, left, right.
cluster_id = math.floor(action / model_action_number)
action_id = action % model_action_number
if action_id == 1:
if self.state_cluster_core_ycor[
cluster_id] + move_step_length_header <= deploy_range_y:
self.state_cluster_core_ycor[
cluster_id] = self.state_cluster_core_ycor[
cluster_id] + move_step_length_header
elif action_id == 2:
if self.state_cluster_core_ycor[
cluster_id] - move_step_length_header >= 0:
self.state_cluster_core_ycor[
cluster_id] = self.state_cluster_core_ycor[
cluster_id] - move_step_length_header
elif action_id == 3:
if self.state_cluster_core_xcor[
cluster_id] + move_step_length_header <= deploy_range_x:
self.state_cluster_core_xcor[
cluster_id] = self.state_cluster_core_xcor[
cluster_id] + move_step_length_header
elif action_id == 4:
if self.state_cluster_core_xcor[
cluster_id] - move_step_length_header >= 0:
self.state_cluster_core_xcor[
cluster_id] = self.state_cluster_core_xcor[
cluster_id] - move_step_length_header
self.set_cluster_id()
self.set_graph()
self.check_graph_error()
def get_reward(self):
reward = self.transmit_flow_in_network()
return reward
# The same as normal DQN procedures.
def choose_action(self, p_state):
if np.random.rand() <= exp_replay.epsilon or len(p_state) == 0:
action = np.random.randint(0, model_output_size)
else:
q = exp_replay.model.predict(p_state)
action = np.argmax(q[0])
return action
# The state in the DQN includes: state_link, cluster_id, and workload.
def get_state(self, tick):
state = []
for i in range(0, len(self.state_link)):
state = state + self.state_link[i]
state = state + self.state_cluster_id
state = state + self.workload_in_network()
state = state + self.event_position
self.check_graph_error()
p_state = np.asarray(state)
p_state = p_state[np.newaxis]
return p_state
def act_action(self, p_state, tick):
if flag_benchmark_action == 1:
action = 0
else:
action = self.choose_action(p_state)
self.action_route(action, tick)
return action
def act_reward(self, tick):
reward = self.get_reward()
p_next_state = self.get_state(tick)
return reward, p_next_state
# Init the event. Theoretically, the event could be init at any places. In the experiment, to make sure that the clusters are somehow balanced, we select a place that is in the middle position of all clusters.
def set_events(self):
if read_AGV:
self.agv = AGV('testing')
self.df = self.agv.get_df()
self.df = self.agv.fit_to_canvas(self.df, deploy_range)
self.df_sampled = self.agv.identical_sample_rate(
self.df, sample_period)
self.grided_pos = self.agv.trace_grided(self.df_sampled, grid_size)
self.trace_grided = self.agv.delete_staying_pos(self.grided_pos)
self.trace_grided = np.asarray(self.trace_grided)
self.trace_input, self.trace_output = self.agv.get_inputs_and_outputs(
self.trace_grided, time_steps)
self.trace_dir_output = self.agv.get_dir(self.trace_input,
self.trace_output)
self.trace_input *= grid_size
self.trace_output *= grid_size
self.trace_grided = self.trace_input[:, :, -1]
self.act_move_events
else:
min_abs_node_id = 0
min_abs_cluster_nodes = total_node_number**2
for i in range(0, total_node_number):
px = self.state_xcor[i]
py = self.state_ycor[i]
cluster_nodes_num_in_range = [0] * server_number
for j in range(0, total_node_number):
if (self.state_xcor[j] -
px)**2 + (self.state_ycor[j] -
py)**2 < event_detection_range_square:
cluster_id = self.state_cluster_id[j]
cluster_nodes_num_in_range[
cluster_id] = cluster_nodes_num_in_range[
cluster_id] + 1
mean_cluster_node_num = np.mean(cluster_nodes_num_in_range)
total_diff_cluster_node_num = 0
for k in range(0, len(cluster_nodes_num_in_range)):
total_diff_cluster_node_num = total_diff_cluster_node_num + abs(
cluster_nodes_num_in_range[k] - mean_cluster_node_num)
if total_diff_cluster_node_num / len(
cluster_nodes_num_in_range) < min_abs_cluster_nodes:
min_abs_cluster_nodes = total_diff_cluster_node_num / len(
cluster_nodes_num_in_range)
min_abs_node_id = i
self.state_event_xcor = self.state_xcor[min_abs_node_id]
self.state_event_ycor = self.state_ycor[min_abs_node_id]
self.event_position = [
self.state_event_xcor, self.state_event_ycor
]
# Select a direction to move the event.
if min_abs_node_id < server_number:
direction_node_id = (min_abs_node_id +
1) % server_number #????
else:
direction_node_id = 1 #????
# Calculate the step distance of the event based on the event speed.
w = (self.state_xcor[direction_node_id] - self.state_event_xcor)
h = (self.state_ycor[direction_node_id] - self.state_event_ycor)
s = (h**2 + w**2)**0.5
self.state_event_step_rate_xcor = w / s
self.state_event_step_rate_ycor = h / s
# The event moves in the network. The event triggers the neighbor sensors to produce more data.
def act_move_events(self):
if read_AGV:
self.state_event_xcor = self.trace_input[total_tick, -1, 0]
self.state_event_ycor = self.trace_input[total_tick, -1, 1]
self.behavior = self.trace_dir_output[total_tick]
self.event_position = [
self.state_event_xcor, self.state_event_ycor
]
else:
self.state_event_xcor = self.state_event_xcor + self.state_event_step_rate_xcor * move_step_length_event
self.state_event_ycor = self.state_event_ycor + self.state_event_step_rate_ycor * move_step_length_event
self.event_position = [
self.state_event_xcor, self.state_event_ycor
]
if self.state_event_xcor < 0 or self.state_event_xcor > deploy_range_x or self.state_event_ycor < 0 or self.state_event_ycor > deploy_range_y:
self.state_event_step_rate_xcor = self.state_event_step_rate_xcor * -1
self.state_event_step_rate_ycor = self.state_event_step_rate_ycor * -1
agv = AGV('training', 0.25)
df = agv.get_df()
df = agv.fit_to_canvas(df, 15)
df_sampled = agv.identical_sample_rate(df, sample_period)
grided_pos = agv.trace_grided(df_sampled, grid_size)
trace_grided = agv.delete_staying_pos(grided_pos)
time_steps = 60
train_propotion = 0.8
train_set_num = int(round(len(trace_grided) * train_propotion))
train_set = trace_grided[:train_set_num]
test_set = trace_grided[train_set_num:]
train_input, train_output = agv.get_inputs_and_outputs(train_set, time_steps)
test_input, test_output = agv.get_inputs_and_outputs(test_set, time_steps)
train_dir_output = agv.get_dir(train_input, train_output)
test_dir_output = agv.get_dir(test_input, test_output)
train_input *= grid_size
train_output *= grid_size
test_input *= grid_size
test_output *= grid_size
# build your neural net
if categorial == True:
model = Sequential([
LSTM(64, input_dim=2),
Activation('relu'),
Dropout(0.15),
Dense(64),
Activation('linear'),