8
LightZero 轻量且高效的 MCTS,AlphaZero,MuZero 系列算法工具包。
-
类别:机器学习
-
项目标题: LightZero:轻量且高效的 MCTS,AlphaZero,MuZero 系列算法工具包。
-
项目描述:LightZero 是一个基于 PyTorch ,集成了蒙特卡洛树搜索(MCTS)和深度强化学习(DRL)的轻量、高效且易于理解的开源算法工具包。 LightZero 致力于推进 MCTS+RL 算法体系的标准化,以便加速相关领域的研究和应用进程,为各种基于 MCTS 的算法和应用提供坚实支持,系统性地支持了下列算法:
- 经典 MCTS 算法
- MCTS+RL 的集大成者 AlphaZero 算法
- 在学习的动力学模型中搜索的 MuZero 算法
- 高样本效率的 EfficientZero 算法
- 基于采样原理可用于多种动作空间的 Sampled MuZero 算法
- 在低搜索开销下也有高性能的 Gumbel MuZero 算法
- 支持随机性环境的 Stochastic MuZero 算法 同时 LightZero 致力于覆盖全面广泛的决策智能基准环境,包括棋盘游戏 (TicTacToe,Gomoku,Chess, Go)、经典控制环境、Atari、MuJoCo、稀疏奖励环境(MiniGrid)、结构化动作空间环境(GoBigger)以及 Bsuite 等。 在 LightZero 的统一的框架下,系统的比较了各种算法,给出了基线结果以及相关分析探究。
-
亮点:
-
示例代码:(可选)
from easydict import EasyDict
# options={'PongNoFrameskip-v4', 'QbertNoFrameskip-v4', 'MsPacmanNoFrameskip-v4', 'SpaceInvadersNoFrameskip-v4', 'BreakoutNoFrameskip-v4', ...}
env_name = 'PongNoFrameskip-v4'
if env_name == 'PongNoFrameskip-v4':
action_space_size = 6
elif env_name == 'QbertNoFrameskip-v4':
action_space_size = 6
elif env_name == 'MsPacmanNoFrameskip-v4':
action_space_size = 9
elif env_name == 'SpaceInvadersNoFrameskip-v4':
action_space_size = 6
elif env_name == 'BreakoutNoFrameskip-v4':
action_space_size = 4
# ==============================================================
# begin of the most frequently changed config specified by the user
# ==============================================================
collector_env_num = 8
n_episode = 8
evaluator_env_num = 3
num_simulations = 50
update_per_collect = 1000
batch_size = 256
max_env_step = int(1e6)
reanalyze_ratio = 0.
# ==============================================================
# end of the most frequently changed config specified by the user
# ==============================================================
atari_muzero_config = dict(
exp_name=
f'data_mz_ctree/{env_name[:-14]}_muzero_ns{num_simulations}_upc{update_per_collect}_rr{reanalyze_ratio}_seed0',
env=dict(
stop_value=int(1e6),
env_name=env_name,
obs_shape=(4, 96, 96),
collector_env_num=collector_env_num,
evaluator_env_num=evaluator_env_num,
n_evaluator_episode=evaluator_env_num,
manager=dict(shared_memory=False, ),
),
policy=dict(
model=dict(
observation_shape=(4, 96, 96),
frame_stack_num=4,
action_space_size=action_space_size,
downsample=True,
self_supervised_learning_loss=True, # default is False
discrete_action_encoding_type='one_hot',
norm_type='BN',
),
cuda=True,
env_type='not_board_games',
game_segment_length=400,
use_augmentation=True,
update_per_collect=update_per_collect,
batch_size=batch_size,
optim_type='SGD',
lr_piecewise_constant_decay=True,
learning_rate=0.2,
num_simulations=num_simulations,
reanalyze_ratio=reanalyze_ratio,
ssl_loss_weight=2, # default is 0
n_episode=n_episode,
eval_freq=int(2e3),
replay_buffer_size=int(1e6), # the size/capacity of replay_buffer, in the terms of transitions.
collector_env_num=collector_env_num,
evaluator_env_num=evaluator_env_num,
),
)
atari_muzero_config = EasyDict(atari_muzero_config)
main_config = atari_muzero_config
atari_muzero_create_config = dict(
env=dict(
type='atari_lightzero',
import_names=['zoo.atari.envs.atari_lightzero_env'],
),
env_manager=dict(type='subprocess'),
policy=dict(
type='muzero',
import_names=['lzero.policy.muzero'],
),
collector=dict(
type='episode_muzero',
import_names=['lzero.worker.muzero_collector'],
)
)
atari_muzero_create_config = EasyDict(atari_muzero_create_config)
create_config = atari_muzero_create_config
if __name__ == "__main__":
from lzero.entry import train_muzero
train_muzero([main_config, create_config], seed=0, max_env_step=max_env_step)
-
截图:(可选)gif/png/jpg
-
后续更新计划:详细的开发计划可以参考 LightZero 的社区 Roadmap