MuZero

class srl.algorithms.muzero.Config(batch_size: int = 32, memory_capacity: int = 100000, memory_warmup_size: int = 1000, memory_compress: bool = True, memory_compress_level: int = -1, observation_mode: str | ~srl.base.define.ObservationModes = ObservationModes.ENV, override_observation_type: ~srl.base.define.SpaceTypes = SpaceTypes.UNKNOWN, override_action_type: str | ~srl.base.define.RLBaseActTypes = <RLBaseActTypes.NONE: 1>, action_division_num: int = 10, observation_division_num: int = 1000, frameskip: int = 0, extend_worker: ~typing.Type[ExtendWorker] | None = None, parameter_path: str = '', memory_path: str = '', use_rl_processor: bool = True, processors: ~typing.List[RLProcessor] = <factory>, render_image_processors: ~typing.List[RLProcessor] = <factory>, enable_state_encode: bool = True, enable_action_decode: bool = True, enable_reward_encode: bool = True, enable_done_encode: bool = True, window_length: int = 1, render_image_window_length: int = 1, enable_sanitize: bool = True, enable_assertion: bool = False, num_simulations: int = 20, discount: float = 0.99, lr: float | ~srl.rl.schedulers.scheduler.SchedulerConfig = 0.001, v_min: int = -10, v_max: int = 10, policy_tau: float | ~srl.rl.schedulers.scheduler.SchedulerConfig = 0.25, unroll_steps: int = 3, root_dirichlet_alpha: float = 0.3, root_exploration_fraction: float = 0.25, c_base: float = 19652, c_init: float = 1.25, dynamics_blocks: int = 15, reward_dense_units: int = 0, weight_decay: float = 0.0001, enable_rescale: bool = True)

<PriorityExperienceReplay> <RLConfigComponentInput>

num_simulations: int = 20

シミュレーション回数

discount: float = 0.99

割引率

lr: float | SchedulerConfig = 0.001

<Scheduler> Learning rate

v_min: int = -10

カテゴリ化する範囲

v_max: int = 10

カテゴリ化する範囲

policy_tau: float | SchedulerConfig = 0.25

policyの温度パラメータのリスト

unroll_steps: int = 3

unroll_steps

root_dirichlet_alpha: float = 0.3

Root prior exploration noise.

root_exploration_fraction: float = 0.25

Root prior exploration noise.

c_base: float = 19652

PUCT

c_init: float = 1.25

PUCT

dynamics_blocks: int = 15

Dynamics networkのブロック数

reward_dense_units: int = 0

reward dense units

weight_decay: float = 0.0001

weight decay

enable_rescale: bool = True

rescale