BitFlip¶
概述¶
BitFlip (翻硬币)是一个非常简单的小游戏。假设有 n 个硬币,每个硬币具有正面和反面两个状态,正面记为 0,反面记为 1。动作空间是一个长度为 n 的向量,执行第 n 个动作类型表示将第 n 个硬币翻转。 对于每一个 epsode,我们随机初始化硬币状态和目标状态。如果硬币状态和目标状态不一样,reward 为 -1,否则为 1。
安装¶
安装方法¶
BitFlip 环境无需安装,内置在 DI-engine 中。
DI-zoo 可运行代码示例¶
下面提供一个完整的 Bitflip 环境 config,采用 DQN 算法作为policy。请在DI-engine/dizoo/classic_control/bitflip/entry
目录下运行bitflip_dqn_main.py
文件,如下。
import os
import gym
from tensorboardX import SummaryWriter
from easydict import EasyDict
from functools import partial
from ding.config import compile_config
from ding.worker import BaseLearner, EpisodeSerialCollector, InteractionSerialEvaluator, EpisodeReplayBuffer
from ding.envs import BaseEnvManager, DingEnvWrapper
from ding.policy import DQNPolicy
from ding.model import DQN
from ding.utils import set_pkg_seed
from ding.rl_utils import get_epsilon_greedy_fn
from ding.reward_model import HerRewardModel
from dizoo.classic_control.bitflip.envs import BitFlipEnv
from dizoo.classic_control.bitflip.config import bitflip_pure_dqn_config, bitflip_her_dqn_config
def main(cfg, seed=0, max_iterations=int(1e8)):
cfg = compile_config(
cfg,
BaseEnvManager,
DQNPolicy,
BaseLearner,
EpisodeSerialCollector,
InteractionSerialEvaluator,
EpisodeReplayBuffer,
save_cfg=True
)
collector_env_num, evaluator_env_num = cfg.env.collector_env_num, cfg.env.evaluator_env_num
collector_env = BaseEnvManager(
env_fn=[partial(BitFlipEnv, cfg=cfg.env) for _ in range(collector_env_num)], cfg=cfg.env.manager
)
evaluator_env = BaseEnvManager(
env_fn=[partial(BitFlipEnv, cfg=cfg.env) for _ in range(evaluator_env_num)], cfg=cfg.env.manager
)
# Set random seed for all package and instance
collector_env.seed(seed)
evaluator_env.seed(seed, dynamic_seed=False)
set_pkg_seed(seed, use_cuda=cfg.policy.cuda)
# Set up RL Policy
model = DQN(**cfg.policy.model)
policy = DQNPolicy(cfg.policy, model=model)
# Set up collection, training and evaluation utilities
tb_logger = SummaryWriter(os.path.join('./{}/log/'.format(cfg.exp_name), 'serial'))
learner = BaseLearner(cfg.policy.learn.learner, policy.learn_mode, tb_logger, exp_name=cfg.exp_name)
collector = EpisodeSerialCollector(
cfg.policy.collect.collector, collector_env, policy.collect_mode, tb_logger, exp_name=cfg.exp_name
)
evaluator = InteractionSerialEvaluator(
cfg.policy.eval.evaluator, evaluator_env, policy.eval_mode, tb_logger, exp_name=cfg.exp_name
)
replay_buffer = EpisodeReplayBuffer(
cfg.policy.other.replay_buffer, exp_name=cfg.exp_name, instance_name='episode_buffer'
)
# Set up other modules, etc. epsilon greedy, hindsight experience replay
eps_cfg = cfg.policy.other.eps
epsilon_greedy = get_epsilon_greedy_fn(eps_cfg.start, eps_cfg.end, eps_cfg.decay, eps_cfg.type)
her_cfg = cfg.policy.other.get('her', None)
if her_cfg is not None:
her_model = HerRewardModel(her_cfg, cfg.policy.cuda)
# Training & Evaluation loop
for _ in range(max_iterations):
# Evaluating at the beginning and with specific frequency
if evaluator.should_eval(learner.train_iter):
stop, reward = evaluator.eval(learner.save_checkpoint, learner.train_iter, collector.envstep)
if stop:
break
# Update other modules
eps = epsilon_greedy(collector.envstep)
# Sampling data from environments
new_episode = collector.collect(train_iter=learner.train_iter, policy_kwargs={'eps': eps})
replay_buffer.push(new_episode, cur_collector_envstep=collector.envstep)
# Training
for i in range(cfg.policy.learn.update_per_collect):
if her_cfg and her_model.episode_size is not None:
sample_size = her_model.episode_size
else:
sample_size = learner.policy.get_attribute('batch_size')
train_episode = replay_buffer.sample(sample_size, learner.train_iter)
if train_episode is None:
break
train_data = []
if her_cfg is not None:
her_episodes = []
for e in train_episode:
her_episodes.extend(her_model.estimate(e))
for e in her_episodes:
train_data.extend(policy.collect_mode.get_train_sample(e))
learner.train(train_data, collector.envstep)
if __name__ == "__main__":
# main(bitflip_pure_dqn_config)
main(bitflip_her_dqn_config)
参考资料¶
BitFlip 源码