Experiment
PPO agents trained in a selfplay settings. This repo includes checkpoints collected during training for 4 experiments:
- Shared weights for actor and critic
- No shared weights
- Resume training for extra steps for both shared and no shared setup Please check our wandb report for more details and the training code on our GitHub
Environment
Multiplayer pong_v3 from PettingZoo with :
- 4 stacked frame
- Agent is trained to predict left agent policy (observation is mirrored for right agent)
def pong_obs_modification(obs, _space, player_id):
obs[:9, :, :] = 0
if "second" in player_id:
# Mirror the image
obs = obs[:, ::-1, :]
return obs
def get_env(args, run_name):
env = importlib.import_module(f"pettingzoo.atari.{args.env_id}").parallel_env()
env = ss.max_observation_v0(env, 2)
env = ss.frame_skip_v0(env, 4)
env = ss.clip_reward_v0(env, lower_bound=-1, upper_bound=1)
env = ss.color_reduction_v0(env, mode="B")
env = ss.resize_v1(env, x_size=84, y_size=84)
env = ss.frame_stack_v1(env, 4)
# Remove the score from the observation
if "pong" in args.env_id:
env = ss.lambda_wrappers.observation_lambda_v0(
env,
pong_obs_modification,
)
# env = ss.agent_indicator_v0(env, type_only=False)
env = ss.pettingzoo_env_to_vec_env_v1(env)
envs = ss.concat_vec_envs_v1(env, args.num_envs // 2, num_cpus=0, base_class="gym")
envs.single_observation_space = envs.observation_space
envs.single_action_space = envs.action_space
envs.is_vector_env = True
envs = gym.wrappers.RecordEpisodeStatistics(envs)
if args.capture_video:
envs = gym.wrappers.RecordVideo(envs, f"videos/{run_name}")
assert isinstance(
envs.single_action_space, gym.spaces.Discrete
), "only discrete action space is supported"
return envs
Model architecture
def atari_network(orth_init=False):
init = layer_init if orth_init else lambda m: m
return nn.Sequential(
init(nn.Conv2d(4, 32, 8, stride=4)),
nn.ReLU(),
init(nn.Conv2d(32, 64, 4, stride=2)),
nn.ReLU(),
init(nn.Conv2d(64, 64, 3, stride=1)),
nn.ReLU(),
nn.Flatten(),
init(nn.Linear(64 * 7 * 7, 512)),
nn.ReLU(),
)
class Agent(nn.Module):
def __init__(self, envs, share_network=False):
super().__init__()
self.actor_network = atari_network(orth_init=True)
self.share_network = share_network
if share_network:
self.critic_network = self.actor_network
else:
self.critic_network = atari_network(orth_init=True)
self.actor = layer_init(nn.Linear(512, envs.single_action_space.n), std=0.01)
self.critic = layer_init(nn.Linear(512, 1), std=1)
def get_value(self, x):
x = x.clone()
x[:, :, :, [0, 1, 2, 3]] /= 255.0
return self.critic(self.critic_network(x.permute((0, 3, 1, 2))))
def get_action_and_value(self, x, action=None):
x = x.clone()
x[:, :, :, [0, 1, 2, 3]] /= 255.0
logits = self.actor(self.actor_network(x.permute((0, 3, 1, 2))))
probs = Categorical(logits=logits)
if action is None:
action = probs.sample()
return (
action,
probs.log_prob(action),
probs.entropy(),
self.critic(self.critic_network(x.permute((0, 3, 1, 2)))),
)
def load(self, path):
self.load_state_dict(torch.load(path))
if self.share_network:
self.critic_network = self.actor_network