def write_train_default_config():
"""
在每次训练之前，修改 default.yaml 文件中的 checkpoint_path 和 cal_max_expectation_tasks 配置项，
修改 qmix.yaml 中的 epsilon_finish 的配置项
:return:
"""
default = ModifyYAML(os.path.join(os.path.dirname(__file__), "config", "default.yaml"))
alg_config = ModifyYAML(os.path.join(os.path.dirname(__file__), "config", "algs", "qmix.yaml"))
default_data = default.data
default_data["checkpoint_path"] = ""
default_data["cal_max_expectation_tasks"] = False
alg_config.data["epsilon_finish"] = 0.1
default.dump()
alg_config.dump()
def write_gen_default_config(checkpoint_path):
"""
在生成数据之前，修改 default.yaml 文件中的 checkpoint_path 和 cal_max_expectation_tasks 配置项，
修改 qmix.yaml 中的 epsilon_finish 的配置项
:param checkpoint_path:
:return:
"""
default = ModifyYAML(os.path.join(os.path.dirname(__file__), "config", "default.yaml"))
alg_config = ModifyYAML(os.path.join(os.path.dirname(__file__), "config", "algs", "qmix.yaml"))
default_data = default.data
default_data["checkpoint_path"] = os.path.join(os.path.dirname(__file__), "results", "models", checkpoint_path)
default_data["cal_max_expectation_tasks"] = True
alg_config.data["epsilon_finish"] = 0
default.dump()
alg_config.dump()
def plot_weight_load():
flag = "heavy_load"
random_mean_reward, all_local_max_expectation, all_offload_max_expectation, rl_max_expectation = get_reward_data(
flag)
print(random_mean_reward.tolist())
print(all_local_max_expectation.tolist())
print(all_offload_max_expectation.tolist())
print(rl_max_expectation.tolist())
x = ModifyYAML(get_ec_config_file_path()).data["light_load_prob"]
png_file_path = os.path.join(os.path.dirname(__file__), "envs", "ec", "output", "heavy_load.png")
eps_file_path = os.path.join(os.path.dirname(__file__), "envs", "ec", "output", "heavy_load.eps")
plot(x, random_mean_reward, all_local_max_expectation, all_offload_max_expectation,
rl_max_expectation, "probability of heavy load", r"normalized $\bar{\psi}$", png_file_path, eps_file_path,
(5.6, 5.6), 16)
def plot_scale():
flag = "cc_cl"
random_mean_reward, all_local_max_expectation, all_offload_max_expectation, rl_max_expectation = get_reward_data(
flag)
print(random_mean_reward.tolist())
print(all_local_max_expectation.tolist())
print(all_offload_max_expectation.tolist())
print(rl_max_expectation.tolist())
x = ModifyYAML(get_ec_config_file_path()).data["cc_cl_scale"]
png_file_path = os.path.join(os.path.dirname(__file__), "envs", "ec", "output", "cc_cl.png")
eps_file_path = os.path.join(os.path.dirname(__file__), "envs", "ec", "output", "cc_cl.eps")
plot(x, random_mean_reward, all_local_max_expectation, all_offload_max_expectation,
rl_max_expectation, r"$C_c / C_l$", r"normalized $\bar{\psi}$", png_file_path, eps_file_path,
(8, 6), 12)
def cal_max_expectation_tasks(args, mac, learner, runner):
"""
加载已经训练好的模型，来生成 state-action-reward 数据，用来评价训练好的模型
:param args:
:param mac:
:param learner:
:param runner:
:return:
"""
algs_modify = ModifyYAML(
os.path.join(os.path.dirname(__file__), "config", "algs", "qmix.yaml"))
algs_modify.data["epsilon_finish"] = 0
algs_modify.dump()
modify = ModifyYAML(
os.path.join(os.path.dirname(__file__), "config", "envs", "ec.yaml"))
global_state = []
global_action = []
global_reward = []
episode = int(modify.data["gen_t_max"] /
modify.data["env_args"]["max_steps"])
for i in range(episode):
episode_batch = runner.run(test_mode=False)
episode_data = episode_batch.data.transition_data
global_state += get_episode_state(episode_data)
global_action += get_episode_action(episode_data)
global_reward += get_episode_reward(episode_data)
expected_reward = ExpectedReward(global_state,
global_reward).get_expected_reward()
label = get_label(modify)
file_path = os.path.join(os.path.dirname(__file__), "envs", "ec", "output",
"rl_" + label + ".txt")
with open(file_path, "a") as f:
f.write(str(expected_reward) + "\n")
all_offload_max_expectation = np.divide(all_offload_max_expectation, max_reward)
rl_max_expectation = np.divide(rl_max_expectation, max_reward)
return random_mean_reward, all_local_max_expectation, all_offload_max_expectation, rl_max_expectation
def my_main():
"""
主函数，系统调用 main.py
:return:
"""
os.system("python3 /home/csyi/pymarl/src/main.py --env-config=ec --config=qmix")
if __name__ == '__main__':
ec_modify = ModifyYAML(os.path.join(os.path.dirname(__file__), "config", "envs", "ec.yaml"))
# 当 train_cc_cl_scale 为 true 时， 表示此时会修改 cc 和 cl 参数，并且训练模型
# 此时必须设置 cc_cl_scale 参数，而且 cc/cl 的比例会依次为 cc_cl_scale 中的值
if ec_modify.data["train_cc_cl_scale"] is True:
cc_cl_scale(ec_modify)
# 当 train_light_load_prob 为 true 时，表示此时会修改 prob 参数，并且训练模型
# 此时必须设置 light_load_prob 参数， 从 light_load_prob 中依次取值来当做轻载的概率
if ec_modify.data["train_light_load_prob"] is True:
light_load_prob(ec_modify)
# 当 train_mid_load_prob 为 true 时，表示此时会修改 prob 参数，并且训练模型
# 此时必须设置 mid_load_prob 参数， 从 mid_load_prob 中依次取值来当做中载的概率
if ec_modify.data["train_mid_load_prob"] is True:
mid_load_prob(ec_modify)