rlu_dmlab_rooms_watermaze

  • 설명 :

RL Unplugged는 오프라인 강화 학습을 위한 벤치마크 모음입니다. RL Unplugged는 다음 고려 사항을 중심으로 설계되었습니다. 사용 편의성을 높이기 위해 일반 파이프라인이 설정되면 실무자가 제품군의 모든 데이터로 쉽게 작업할 수 있도록 하는 통합 API와 함께 데이터 세트를 제공합니다.

데이터 세트는 RLDS 형식 을 따라 단계와 에피소드를 나타냅니다.

DeepMind Lab 데이터 세트에는 도전적이고 부분적으로 관찰 가능한 Deepmind Lab 제품군 의 여러 수준이 있습니다. DeepMind Lab 데이터 세트는 Kapturowski et al., 2018 에이전트가 개별 작업에 대해 처음부터 분산 R2D2를 교육하여 수집됩니다. 모든 작업에 대해 전체 교육 실행 동안 모든 행위자에 대한 경험을 몇 번 기록했습니다. 데이터 세트 생성 프로세스의 세부 사항은 Gulcehre et al., 2021 에 설명되어 있습니다.

5가지 DeepMind Lab 레벨( seekavoid_arena_01 , explore_rewards_few , explore_rewards_many , rooms_watermaze , rooms_select_nonmatching_object )에 대한 데이터 세트를 공개합니다. 또한 환경에서 에이전트를 평가할 때 epsilon-greedy 알고리즘에 대해 서로 다른 수준의 엡실론을 사용하여 훈련된 R2D2 스냅샷에서 데이터 세트를 생성한 seekavoid_arena_01 수준의 스냅샷 데이터 세트를 릴리스합니다.

DeepMind Lab 데이터 세트는 상당히 대규모입니다. 메모리가 포함된 대규모 오프라인 RL 모델에 관심이 있는 경우 사용해 보는 것이 좋습니다.

FeaturesDict({
    'episode_id': int64,
    'episode_return': float32,
    'steps': Dataset({
        'action': int64,
        'discount': float32,
        'is_first': bool,
        'is_last': bool,
        'is_terminal': bool,
        'observation': FeaturesDict({
            'last_action': int64,
            'last_reward': float32,
            'pixels': Image(shape=(72, 96, 3), dtype=uint8),
        }),
        'reward': float32,
    }),
})
  • 기능 문서 :
특징 수업 모양 D타입 설명
풍모Dict
episode_id 텐서 int64
episode_return 텐서 float32
단계 데이터 세트
단계/액션 텐서 int64
단계/할인 텐서 float32
단계/is_first 텐서 부울
단계/is_last 텐서 부울
단계/is_terminal 텐서 부울
단계/관찰 풍모Dict
단계/관찰/last_action 텐서 int64
단계/관찰/last_reward 텐서 float32
단계/관찰/픽셀 영상 (72, 96, 3) uint8
단계/보상 텐서 float32
@article{gulcehre2021rbve,
    title={Regularized Behavior Value Estimation},
    author={ {\c{C} }aglar G{\"{u} }l{\c{c} }ehre and
               Sergio G{\'{o} }mez Colmenarejo and
               Ziyu Wang and
               Jakub Sygnowski and
               Thomas Paine and
               Konrad Zolna and
               Yutian Chen and
               Matthew W. Hoffman and
               Razvan Pascanu and
               Nando de Freitas},
    year={2021},
    journal   = {CoRR},
    url       = {https://arxiv.org/abs/2103.09575},
    eprint={2103.09575},
    archivePrefix={arXiv},
}

rlu_dmlab_rooms_watermaze/training_0(기본 구성)

  • 데이터세트 크기 : 894.50 GiB

  • 분할 :

나뉘다
'train' 67,876

rlu_dmlab_rooms_watermaze/training_1

  • 데이터세트 크기 : 898.74 GiB

  • 분할 :

나뉘다
'train' 66,922

rlu_dmlab_rooms_watermaze/training_2

  • 데이터세트 크기 : 825.49 GiB

  • 분할 :

나뉘다
'train' 67,081