def ppo_continuous(game, tag=""):
config = PPOConfig()
config.num_workers = 16
config.task_fn = lambda: Task(
game, num_envs=config.num_workers, single_process=False)
config.eval_env = Task(game)
config.optimizer_fn = lambda params: Adam(params, 3e-4, eps=1e-5)
config.network_fn = lambda: GaussianActorCriticNet(
config.state_dim,
config.action_dim,
actor_body=FCBody(config.state_dim, gate=F.tanh),
critic_body=FCBody(config.state_dim, gate=F.tanh))
# config.state_normalizer =Mea
config.discount = 0.99
config.use_gae = True
config.gae_tau = 0.95
config.entropy_weight = 0.01
config.rollout_length = 2048
config.gradient_clip = 0.5
config.optimization_epochs = 10
config.mini_batch_size = 64
config.ppo_ratio_clip = 0.2
config.log_interval = 2048
config.max_steps = int(1e6)
PPOAgent(config).run_steps(tag=f'{tag}{ppo_continuous.__name__}-{game}')
def categorical_dqn_cart_pole():
game = 'CartPole-v0'
config = CategoricalDQNConfig()
config.task_fn = lambda: Task(game)
config.eval_env = Task(game)
config.optimizer_fn = lambda params: RMSprop(params, 0.001)
config.network_fn = lambda: CategoricalNet(config.action_dim, config.
categorical_n_atoms,
FCBody(config.state_dim))
config.batch_size = 10
config.replay_fn = lambda: ReplayBuffer(config.eval_env,
memory_size=int(1e4))
config.random_action_prob = LinearSchedule(1.0, 0.1, 1e4)
config.discount = 0.99
config.target_network_update_freq = 200
config.exploration_steps = 100
config.categorical_v_max = 100
config.categorical_v_min = -100
config.categorical_n_atoms = 50
config.rollout_length = 4
config.gradient_clip = 5
config.max_steps = 1e6
CategoricalDQNAgent(config).run_steps(
tag=f'{categorical_dqn_cart_pole.__name__}-{game}')
def ddpg_continuous(game, tag=""):
config = DDPGConfig()
config.task_fn = lambda: Task(game)
config.eval_env = Task(game)
config.network_fn = lambda: DeterministicActorCriticNet(
config.state_dim,
config.action_dim,
actor_body=FCBody(config.state_dim, (400, 300), gate=F.relu),
critic_body=TwoLayerFCBodyWithAction(
config.state_dim, config.action_dim, (400, 300), gate=F.relu),
actor_opt_fn=lambda params: Adam(params, lr=1e-4),
critic_opt_fn=lambda params: Adam(params, lr=1e-3))
config.batch_size = 64
config.replay_fn = lambda: ReplayBuffer(config.eval_env,
memory_size=int(1e6))
config.random_process_fn = lambda: OrnsteinUhlenbeckProcess(
size=(config.action_dim, ), std=LinearSchedule(0.2))
config.discount = 0.99
config.min_memory_size = 64
config.target_network_mix = 1e-3
config.max_steps = int(1e6)
DDPGAgent(config).run_steps(tag=f'{tag}{ddpg_continuous.__name__}-{game}')
def a2c_continuous(game, tag=""):
config = A2CConfig()
config.num_workers = 16
config.task_fn = lambda: Task(
game, num_envs=config.num_workers, single_process=True)
config.eval_env = Task(game)
config.optimizer_fn = lambda params: RMSprop(params, lr=0.0007)
config.network_fn = lambda: GaussianActorCriticNet(
config.state_dim,
config.action_dim,
actor_body=FCBody(config.state_dim),
critic_body=FCBody(config.state_dim))
config.discount = 0.99
config.use_gae = True
config.gae_tau = 1.0
config.entropy_weight = 0.01
config.rollout_length = 5
config.gradient_clip = 5
config.max_steps = int(1e6)
A2CAgent(config).run_steps(tag=f'{tag}{a2c_continuous.__name__}-{game}')
def vpg_cart_pole(game):
config = VPGConfig()
config.num_workers = 5
config.task_fn = lambda: Task(game, num_envs=config.num_workers)
config.eval_env = Task(game)
config.optimizer_fn = lambda params: Adam(params, lr=1e-3)
config.network_fn = lambda: CategoricalActorCriticNet(
config.state_dim, config.action_dim, FCBody(config.state_dim))
config.discount = 0.99
config.use_gae = True
config.gae_tau = 0.97
config.entropy_weight = 0.001
config.rollout_length = 4000
config.gradient_clip = 5
config.logger = get_logger(tag=vpg_cart_pole.__name__)
run_steps(VPGAgent(config))
def option_critic_cart_pole():
game = 'CartPole-v0'
config = OptionCriticConfig()
config.num_workers = 8
config.task_fn = lambda: Task(game, num_envs=config.num_workers, single_process=True)
config.eval_env = Task(game)
config.optimizer_fn = lambda params: RMSprop(params, 0.001)
config.network_fn = lambda: OptionCriticNet(FCBody(config.state_dim), config.action_dim, num_options=2)
config.random_option_prob = LinearSchedule(1.0, 0.01, 1e4)
config.discount = 0.99
config.target_network_update_freq = 200
config.rollout_length = 5
config.termination_regularizer = 0.01
config.entropy_weight = 0.01
config.gradient_clip = 5
config.max_steps = 1e6
OptionCriticAgent(config).run_steps(tag=f'{option_critic_cart_pole.__name__}-{game}')
def a2c_cart_pole():
game = 'CartPole-v0'
config = A2CConfig()
config.num_workers = 16
config.task_fn = lambda: Task(
game, num_envs=config.num_workers, single_process=True)
config.eval_env = Task(game)
config.optimizer_fn = lambda params: Adam(params, lr=1e-3)
config.network_fn = lambda: CategoricalActorCriticNet(
config.state_dim, config.action_dim, FCBody(config.state_dim))
config.discount = 0.99
config.use_gae = True
config.gae_tau = 0.95
config.entropy_weight = 0.01
config.rollout_length = 5
config.gradient_clip = 0.5
config.max_steps = 1e6
A2CAgent(config).run_steps(tag=f'{a2c_cart_pole.__name__}-{game}')
def nstepdqn_cart_pole():
game = 'CartPole-v0'
config = NStepDQNConfig()
config.num_workers = 16
config.task_fn = lambda: Task(
game, num_envs=config.num_workers, single_process=True)
config.eval_env = Task(game)
config.optimizer_fn = lambda params: RMSprop(params, 0.001)
config.network_fn = lambda: VanillaNet(config.action_dim,
FCBody(config.state_dim))
# config.network_fn = lambda: DuelingNet(config.action_dim, FCBody(config.state_dim))
config.random_action_prob = LinearSchedule(1.0, 0.1, 1e4)
config.discount = 0.99
config.target_network_update_freq = 200
config.double_q = True
config.rollout_length = 1
config.gradient_clip = 5
config.max_steps = 1e6
NStepDQNAgent(config).run_steps(
tag=f'{nstepdqn_pixel_atari.__name__}-{game}')
def dqn_cart_pole():
game = 'CartPole-v0'
config = DQNConfig()
config.task_fn = lambda: Task(game)
config.eval_env = Task(game)
config.optimizer_fn = lambda params: RMSprop(params, 0.001)
config.network_fn = lambda: VanillaNet(config.action_dim, FCBody(config.state_dim))
# config.network_fn = lambda: DuelingNet(config.action_dim, FCBody(config.state_dim))
config.batch_size = 10
config.replay_fn = lambda: ReplayBuffer(config.eval_env, memory_size=int(1e4))
config.random_action_prob = LinearSchedule(1.0, 0.1, 1e4)
config.discount = 0.99
config.target_network_update_freq = 200
config.exploration_steps = 1000
config.double_q = True
config.rollout_length = 4
config.gradient_clip = 5
config.eval_interval = int(5e3)
config.max_steps = 1e6
DQNAgent(config).run_steps(tag=f'{dqn_cart_pole.__name__}-{game}')
def ppo_cart_pole():
game = 'CartPole-v0'
config = PPOConfig()
config.num_workers = 8
config.task_fn = lambda: Task(
game, num_envs=config.num_workers, single_process=True)
config.eval_env = Task(game)
config.optimizer_fn = lambda params: RMSprop(params, 0.001)
config.network_fn = lambda: CategoricalActorCriticNet(
config.state_dim, config.action_dim, FCBody(config.state_dim))
config.discount = 0.99
config.use_gae = True
config.gae_tau = 0.95
config.entropy_weight = 0.001
config.gradient_clip = 5
config.rollout_length = 128
config.optimization_epochs = 10
config.mini_batch_size = 32 * 5
config.ppo_ratio_clip = 0.2
config.max_steps = 1e6
PPOAgent(config).run_steps(tag=f'{ppo_cart_pole.__name__}-{game}')