configs.py

ATARI_DEFAULTS = dict(
    # General
    steps=2e8,
    eval_every=1e5,
    log_every=1e4,
    prefill=5e4,
    dataset_size=4e6,
    pretrain=0,

    # Environment
    time_limit=108000,  # 30 minutes of game play.
    action_repeat=4,
    eval_noise=0.001,
    train_every=200,
    train_steps=10,
    clip_rewards='tanh',

    # Model
    pred_discount=True,
    cnn_depth=48,  # 32
    stoch_size=100,
    deter_size=600,

    # Behavior
    actor_dist='onehot',
    actor_entropy=5e-3,
    expl='epsilon_greedy',
    expl_amount=0.0,
    expl_until=5e7,
    discount=0.995,

    # Training
    model_lr=1e-4,  # 2e-4
    actor_lr=4e-5,  # 5e-5
    value_lr=8e-5,  # 1e-4
    weight_decay=0.0,  # 2e-5
    kl_scale=0.3,
    free_nats=0.0,
    actor_grad_clip=5.0,
    value_grad_clip=5.0,
    oversample_ends=True,
    slow_value_target=True,
    slow_actor_target=True,
    slow_target_update=10,
    # Empowerment
)

MINERL_DEFAULTS = dict(
    # General
    steps=6e8,
    eval_every=1e5,
    log_every=1e4,
    prefill=1e4,
    dataset_size=4e6,
    pretrain=0,

    # Environment
    time_limit=8000,
    train_every=200,
    train_steps=10,

    # Model
    pred_discount=True,
    cnn_depth=48,  # 32
    stoch_size=100,
    deter_size=600,
    proprio=True,

    # Behavior
    actor_dist='onehot',
    expl='epsilon_greedy',
    expl_amount=0.0,
    expl_until=5e7,
    discount=0.99,

    # Training
    model_lr=1e-4,  # 2e-4
    actor_lr=5e-5,  # 5e-5
    value_lr=5e-5,  # 1e-4
    weight_decay=0.0,  # 2e-5
    free_nats=3.0,
    actor_grad_clip=100.0,
    value_grad_clip=100.0,
    oversample_ends=True,
    slow_value_target=True,
    slow_actor_target=True,
    slow_target_update=10,
)

DMC_TOY_DEFAULTS = dict(
    # General
    steps=5e6,
    eval_every=1e5,
    log_every=1e4,
    prefill=5e3,
    dataset_size=4e6,
    pretrain=0,

    # Environment
    time_limit=1000,
    eval_noise=0.001,
    train_every=200,
    train_steps=10,

    # Training
    oversample_ends=True,
    slow_value_target=True,
    slow_actor_target=True,
    slow_target_update=10,
)

DMC_DEFAULTS = dict(
    # General
    steps=5e6,
    eval_every=1e5,
    log_every=1e4,
    prefill=5e3,
    dataset_size=4e6,
    pretrain=0,

    # Training
    slow_value_target=True,
    slow_actor_target=True,
    slow_target_update=100,
)

default_configs=dict(atari=ATARI_DEFAULTS, 
                     mc=MINERL_DEFAULTS,
                     dmctoy=DMC_TOY_DEFAULTS,
                     dmc=DMC_DEFAULTS)