def train_fixed(seed, agent_setting, game_name="ma_softq"):
set_seed(seed)
suffix = f"fixed_play1/{game_name}/{agent_setting}/{seed}"
set_logger(suffix)
batch_size = 512
training_steps = 2000
exploration_steps = 100
max_replay_buffer_size = 1e5
hidden_layer_sizes = (128, 128)
max_path_length = 1
agent_num = 2
env = DifferentialGame(game_name, agent_num)
agents = []
agent_types = agent_setting.split("_")
assert len(agent_types) == agent_num
for i, agent_type in enumerate(agent_types):
agents.append(
get_agent_by_type(
agent_type,
i,
env,
hidden_layer_sizes=hidden_layer_sizes,
max_replay_buffer_size=max_replay_buffer_size,
)
)
sampler = MASampler(
agent_num, batch_size=batch_size, max_path_length=max_path_length
)
sampler.initialize(env, agents)
trainer = MATrainer(
env=env,
agents=agents,
sampler=sampler,
steps=training_steps,
exploration_steps=exploration_steps,
training_interval=1,
extra_experiences=["annealing", "recent_experiences"],
batch_size=batch_size,
)
trainer.run()
def train_fixed(seed, agent_setting, game_name='ma_softq'):
set_seed(seed)
suffix = f'fixed_play/{game_name}/{agent_setting}/{seed}'
set_logger(suffix)
batch_size = 1024
training_steps = 25 * 60000
exploration_steps = 2000
max_replay_buffer_size = 1e5
hidden_layer_sizes = (100, 100)
max_path_length = 25
agent_num = 3
env = make_particle_env(game_name)
agents = []
agent_types = agent_setting.split('_')
assert len(agent_types) == agent_num
for i, agent_type in enumerate(agent_types):
agents.append(
get_agent_by_type(agent_type,
i,
env,
hidden_layer_sizes=hidden_layer_sizes,
max_replay_buffer_size=max_replay_buffer_size))
sampler = MASampler(agent_num,
batch_size=batch_size,
max_path_length=max_path_length)
sampler.initialize(env, agents)
trainer = MATrainer(env=env,
agents=agents,
sampler=sampler,
steps=training_steps,
exploration_steps=exploration_steps,
training_interval=10,
extra_experiences=['annealing', 'recent_experiences'],
batch_size=batch_size)
trainer.run()
def train_fixed(seed, agent_setting, env_configs, fullly_centralized):
set_seed(seed)
scenario = env_configs["scenario"]
suffix = f"fixed_play/{scenario}/{agent_setting}/{seed}"
set_logger(suffix)
batch_size = 50
training_steps = 25 * 60000
exploration_steps = 100
max_replay_buffer_size = 1e5
hidden_layer_sizes = (100, 100)
max_path_length = 25
agent_num = env_configs["n_agents"]
raw_env = ListHiWayEnv(env_configs)
env = Wrapper(env=raw_env, action_space=raw_env.action_sapce)
if fullly_centralized:
agent = get_agent_by_type(
agent_setting,
0,
env,
hidden_layer_sizes=hidden_layer_sizes,
max_replay_buffer_size=max_replay_buffer_size,
)
sampler = SingleSampler(batch_size=batch_size,
max_path_length=max_path_length)
sampler.initialize(env, agent)
extra_experiences = ["target_actions"]
trainer = SATrainer(
env=env,
agent=agent,
sampler=sampler,
steps=training_steps,
exploration_steps=exploration_steps,
training_interval=10,
extra_experiences=extra_experiences,
batch_size=batch_size,
)
else:
agents = []
for i in range(agent_num):
agents.append(
get_agent_by_type(
agent_setting,
i,
env,
hidden_layer_sizes=hidden_layer_sizes,
max_replay_buffer_size=max_replay_buffer_size,
))
sampler = MASampler(agent_num,
batch_size=batch_size,
max_path_length=max_path_length)
sampler.initialize(env, agents)
extra_experiences = [
"annealing", "recent_experiences", "target_actions"
]
trainer = MATrainer(
env=env,
agents=agents,
sampler=sampler,
steps=training_steps,
exploration_steps=exploration_steps,
training_interval=10,
extra_experiences=extra_experiences,
batch_size=batch_size,
)
trainer.run()
batch_size = 128
training_steps = 10000
exploration_step = 1000
hidden_layer_sizes = (10, 10)
max_replay_buffer_size = 1e5
env = DifferentialGame(game_name, agent_num)
agents = []
for i in range(agent_num):
agent = get_maddpg_agent(
env,
i,
hidden_layer_sizes=hidden_layer_sizes,
max_replay_buffer_size=max_replay_buffer_size,
)
agents.append(agent)
sampler = MASampler(agent_num)
sampler.initialize(env, agents)
trainer = MATrainer(
env=env,
agents=agents,
sampler=sampler,
steps=training_steps,
exploration_steps=exploration_step,
extra_experiences=["target_actions"],
)
trainer.run()