def Benchmark_test_time(max_episode, Ns, fd, max_dis, maxM):
env = Env_cellular(fd, Ts, n_x, n_y, L, C, maxM, min_dis, max_dis, max_p,
p_n, power_num)
bench = Benchmark_alg(env)
env.set_Ns(Ns)
time_cost_1 = 0
time_cost_2 = 0
for k in range(1, max_episode + 1):
s = env.reset__()
for i in range(int(Ns) - 1):
tc, p = bench.time_cost(s)
time_cost_1 = time_cost_1 + tc[0]
time_cost_2 = time_cost_2 + tc[1]
s_next, _ = env.step__(p)
s = s_next
return [time_cost_1, time_cost_2]
def Benchmark_test(max_episode, Ns, fd, max_dis, maxM):
env = Env_cellular(fd, Ts, n_x, n_y, L, C, maxM, min_dis, max_dis, max_p,
p_n, power_num)
bench = Benchmark_alg(env)
env.set_Ns(Ns)
reward_hist = list()
for k in range(1, max_episode + 1):
reward_list = list()
s = env.reset__()
for i in range(int(Ns) - 1):
p = bench.calculate(s)
s_next, r = env.step__(p)
s = s_next
reward_list.append(r)
reward_hist.append(reward_list)
reward_hist = np.reshape(reward_hist, [max_episode, Ns - 1, 4])
reward_hist = np.nanmean(np.nanmean(reward_hist, 0), 0)
return reward_hist
def Test_policy_time(weight_file, max_episode, Ns, fd, max_dis, maxM):
env = Env_cellular(fd, Ts, n_x, n_y, L, C, maxM, min_dis, max_dis, max_p,
p_n, power_num)
tf.reset_default_graph()
with tf.Session() as sess:
env.set_Ns(Ns)
dnn = DNN(env, weight_file)
policy = Policy(sess, dnn)
tf.global_variables_initializer().run()
policy.load_params()
time_cost = 0
for k in range(1, max_episode + 1):
s_actor, _ = env.reset()
for i in range(int(Ns) - 1):
st = time.time()
p = policy.predict_p(s_actor)
time_cost = time_cost + time.time() - st
s_actor, _, _, _ = env.step(p)
return time_cost
def Test_policy_all(weight_file, max_episode, Ns, fd, max_dis, maxM):
env = Env_cellular(fd, Ts, n_x, n_y, L, C, maxM, min_dis, max_dis, max_p,
p_n, power_num)
tf.reset_default_graph()
with tf.Session() as sess:
env.set_Ns(Ns)
dnn = DNN(env, weight_file)
policy = Policy(sess, dnn)
tf.global_variables_initializer().run()
policy.load_params()
reward_hist = list()
for k in range(1, max_episode + 1):
reward_policy_list = list()
s_actor, _ = env.reset()
for i in range(int(Ns) - 1):
p = policy.predict_p(s_actor)
s_actor, _, _, r = env.step(p)
reward_policy_list.append(r)
reward_hist.append(np.mean(reward_policy_list)) # bps/Hz per link
return np.mean(reward_hist)
def Test_ddpg_all(weight_file, max_episode, Ns, fd, max_dis, maxM):
env = Env_cellular(fd, Ts, n_x, n_y, L, C, maxM, min_dis, max_dis, max_p,
p_n, power_num)
tf.reset_default_graph()
with tf.Session() as sess:
env.set_Ns(Ns)
dnn = DDPG(env, weight_file)
actor = Actor(sess, dnn)
tf.global_variables_initializer().run()
actor.load_params()
reward_hist = list()
for k in range(1, max_episode + 1):
reward_list = list()
s_actor, _ = env.reset()
for i in range(int(Ns) - 1):
p = actor.predict_p(s_actor)
s_actor_next, _, _, r = env.step(p[:, 0])
s_actor = s_actor_next
reward_list.append(r)
reward_hist.append(np.mean(reward_list))
return np.mean(reward_hist)
def Train_policy_quan(weight_file, power_num, fd, max_dis, maxM):
env = Env_cellular(fd, Ts, n_x, n_y, L, C, maxM, min_dis, max_dis, max_p,
p_n, power_num)
tf.reset_default_graph()
with tf.Session() as sess:
Train(sess, env, weight_file)
tf.reset_default_graph()
with tf.Session() as sess:
env.set_Ns(Ns)
dnn = DNN(env, weight_file)
policy = Policy(sess, dnn)
tf.global_variables_initializer().run()
policy.load_params()
time_cost = 0
for k in range(1, max_episode + 1):
s_actor, _ = env.reset()
for i in range(int(Ns) - 1):
st = time.time()
p = policy.predict_p(s_actor)
time_cost = time_cost + time.time() - st
s_actor, _, _, _ = env.step(p)
return time_cost
if __name__ == "__main__":
env = Env_cellular(fd, Ts, n_x, n_y, L, C, maxM, min_dis, max_dis, max_p,
p_n, power_num)
weight_file = './saved_networks/reinforce_2.mat'
tf.reset_default_graph()
with tf.Session() as sess:
train_hist = Train(sess, env, weight_file)
tf.reset_default_graph()
with tf.Session() as sess:
test_hist = Test(sess, env, weight_file)