scenarios = [
'simple_adversary', 'simple_crypto', 'simple_push', 'simple_tag',
'simple_world_comm'
]
TEST_ONLY = False
if TEST_ONLY:
arglist.is_training = False
for scenario_name in scenarios:
arglist.actor_learning_rate = 1e-2
arglist.critic_learning_rate = 1e-2
for cnt in range(10):
# scenario_name = 'simple_spread'
env = make_env(scenario_name, discrete_action=True)
seed = cnt + 12345678
env.seed(seed)
torch.cuda.empty_cache()
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
dim_obs_own = env.observation_space[-1].shape[0]
dim_obs_adv = env.observation_space[0].shape[0]
dim_action_own = env.action_space[-1].n
dim_action_adv = env.action_space[0].n
action_type = 'Discrete'
# num_adv & adv action dims
if __name__ == '__main__':
from rls.model.ac_network_model_multi_gumbel import ActorNetwork, CriticNetwork
from rls.agent.multiagent.model_ddpg_gumbel_fix import Trainer
from experiments.scenarios import make_env
import os
arglist.actor_learning_rate = 1e-2
arglist.critic_learning_rate = 1e-2
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" # see issue #152
os.environ["CUDA_VISIBLE_DEVICES"] = '1'
cnt = 11
# scenario_name = 'simple_spread'
scenario_name = 'simple_speaker_listener'
env = make_env(scenario_name, benchmark=False, discrete_action=True)
seed = cnt + 12345678
env.seed(seed)
torch.cuda.empty_cache()
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
dim_obs = env.observation_space[0].shape[0]
if hasattr(env.action_space[0], 'high'):
dim_action = env.action_space[0].high + 1
dim_action = dim_action.tolist()
action_type = 'MultiDiscrete'
else:
dim_action = env.action_space[0].n
action_type = 'Discrete'
'simple_spread', 'simple_reference', 'simple_speaker_listener',
'fullobs_collect_treasure', 'multi_speaker_listener'
]
TEST_ONLY = False
scenario_name = 'simple_spread'
for n_agent in [6, 9, 12]:
arglist.actor_learning_rate = 1e-2
arglist.critic_learning_rate = 1e-2
for cnt in range(5):
# scenario_name = 'simple_spread'
env = make_env(scenario_name,
n=n_agent,
benchmark=False,
discrete_action=True,
local_observation=True)
seed = cnt + 12345678
# print(env.observation_space)
env.seed(seed)
torch.cuda.empty_cache()
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
dim_obs = env.observation_space[0].shape[0]
if hasattr(env.action_space[0], 'high'):
dim_action = env.action_space[0].high + 1