class ProcessAgent(Process):
def __init__(self, id, prediction_q, training_q):
super(ProcessAgent, self).__init__()
self.id = id
self.prediction_q = prediction_q
self.training_q = training_q
self.env = Environment()
self.wait_q = Queue(maxsize=1)
self.exit_flag = Value('i', 0)
def predict(self, state, current_location, depot_idx):
self.prediction_q.put((self.id, state, current_location, depot_idx))
a, v = self.wait_q.get()
return a, v
def run_episode(self):
if Config.FROM_FILE == 1:
file_state = np.load('data_state_00.npy', 'r')
file_or_route = np.load('data_or_route_00.npy', 'r')
file_or_cost = np.load('data_or_cost_00.npy', 'r')
self.batch_idx = np.random.randint(10000)
self.env.current_state = file_state[self.batch_idx, :]
# print(self.env.current_state)
self.env.distance_matrix = self.env.get_distance_matrix()
current_location = self.env.get_current_location(
) # may need to change in future
idx = int(self.env.get_depot_idx())
or_route = file_or_route[self.batch_idx]
# tmp_or_model = OR_Tool(self.env.current_state, current_location, idx)
# tmp_or_route, tmp_or_cost = tmp_or_model.solve()
# print(or_route, tmp_or_route)
or_cost = file_or_cost[self.batch_idx]
# action = np.zeros([Config.NUM_OF_CUSTOMERS+1], dtype=np.int32)
# base_line = 1.0
# sampled_value = 1.0
action, base_line = self.predict(self.env.current_state,
current_location, idx)
sampled_value = self.env.G(action, current_location)
else:
self.env.reset()
current_location = self.env.get_current_location()
idx = int(self.env.get_depot_idx())
action, base_line = self.predict(self.env.current_state,
current_location, idx)
sampled_value = self.env.G(action, current_location)
if Config.REINFORCE == 0:
or_model = OR_Tool(self.env.current_state, current_location,
idx)
or_route, or_cost = or_model.solve()
else:
or_route = np.zeros([Config.NUM_OF_CUSTOMERS + 1],
dtype=np.int32)
or_cost = 1.0
return action, base_line, sampled_value, or_route, or_cost, idx
def run(self):
np.random.seed(np.int32(time.time() % 1 * 1000 + self.id * 10))
while self.exit_flag.value == 0:
a_, b_, r_, ora_, orr_, idx_ = self.run_episode()
x_ = self.env.current_state
y_ = self.env.get_current_location()
self.training_q.put(
([x_], [y_], [a_], [ora_], [r_], [orr_], [idx_]))
