![今回取り上げるのはこれ
[1] Volodymyr Mnih, Adria` Puigdome`nech Badia, Mehdi
Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David
Silver, and Koray Kavukcuoglu. Asynchronous methods for
deep reinforcement learning. In Proceedings of the 33rd
International Conference on Machine Learning (ICML), pp.
1928–1937, 2016.
Asynchronousな手法によりreplay memoryを廃し、DQNより
高速かつ高精度な学習を達成した！](https://image.slidesharecdn.com/a3cpdf-170115092158/85/Introduction-to-A3C-model-2-320.jpg)
![[DL輪読会]Decision Transformer: Reinforcement Learning via Sequence Modeling](https://cdn.slidesharecdn.com/ss_thumbnails/decisiontransformer20210709zhangxin-210709021501-thumbnail.jpg?width=560&fit=bounds)
![SSII2021 [TS2] 深層強化学習 〜 強化学習の基礎から応用まで 〜](https://cdn.slidesharecdn.com/ss_thumbnails/ts2-01-210607042910-thumbnail.jpg?width=560&fit=bounds)
![[DL輪読会]深層強化学習はなぜ難しいのか？Why Deep RL fails? A brief survey of recent works.](https://cdn.slidesharecdn.com/ss_thumbnails/20210115dlohta-210115054939-thumbnail.jpg?width=560&fit=bounds)
![[DL輪読会]Learning Latent Dynamics for Planning from Pixels](https://cdn.slidesharecdn.com/ss_thumbnails/taniguchi20181221-190104064850-thumbnail.jpg?width=560&fit=bounds)
![[DL輪読会]`強化学習のための状態表現学習 －より良い「世界モデル」の獲得に向けて－](https://cdn.slidesharecdn.com/ss_thumbnails/20181026staterepresenration-181127055206-thumbnail.jpg?width=560&fit=bounds)
![SSII2021 [OS2-01] 転移学習の基礎：異なるタスクの知識を利用するための機械学習の方法](https://cdn.slidesharecdn.com/ss_thumbnails/os2-02final-210610091211-thumbnail.jpg?width=560&fit=bounds)
![[DL輪読会]Set Transformer: A Framework for Attention-based Permutation-Invariant...](https://cdn.slidesharecdn.com/ss_thumbnails/20200221settransformer-200221020423-thumbnail.jpg?width=560&fit=bounds)
![[DL輪読会]Dream to Control: Learning Behaviors by Latent Imagination](https://cdn.slidesharecdn.com/ss_thumbnails/20200313furutav2-200313025657-thumbnail.jpg?width=560&fit=bounds)
Ad
More Related Content
What's hot (20)
Viewers also liked (11)
![[Dl輪読会]introduction of reinforcement learning](https://cdn.slidesharecdn.com/ss_thumbnails/dlintroductionofreinforcementlearning-161121061444-thumbnail.jpg?width=560&fit=bounds)
Ad
Similar to Introduction to A3C model (20)
![[GTCJ2018] Optimizing Deep Learning with Chainer PFN得居誠也](https://cdn.slidesharecdn.com/ss_thumbnails/gtcj2018optimizingdeeplearningwithpfnseiyatokui-181009073509-thumbnail.jpg?width=560&fit=bounds)
Ad
More from WEBFARMER. ltd. (20)
Introduction to A3C model
- 2. 今回取り上げるのはこれ [1] Volodymyr Mnih, Adria` Puigdome`nech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Proceedings of the 33rd International Conference on Machine Learning (ICML), pp. 1928–1937, 2016. Asynchronousな手法によりreplay memoryを廃し、DQNより 高速かつ高精度な学習を達成した！
- 4. 強化学習の基本① Li θi( )= E r +γ max a' Q s',a';θi−1( )−Q s,a;θi( )( ) 2 1-step Q学習の損失関数 actor-criticにおける 目的関数の勾配 1-step Sarsaの損失関数 Li θi( )= E r +γQ s',a';θi−1( )−Q s,a;θi( )( ) 2 n-step Q学習の損失関数 Li θi( )= E γk rt+k k=0 n ∑ + maxγ a' n Q s',a';θi−1( )−Q s,a;θi( ) ⎛ ⎝ ⎜ ⎞ ⎠ ⎟ 2 ∇θ J θ( )= E ∇θ logπ at | st;θ( ) Rt −Vπ st( )( )⎡ ⎣ ⎤ ⎦ r γ Q s,a;θi( ) Vπ st( ) ：割引率 ：報酬 ：状態 s で行動 a を取る場合の行動価値関数 ：状態 s の価値関数
- 5. 強化学習の基本② Li θi( )= E r +γ max a' Q s',a';θi−1( )−Q s,a;θi( )( ) 2 1-step Q学習の損失関数 これがDQNの場合 L θ( )= Es,a,r,s'≈D r +γ max a' Q s',a';θ− ( )−Q s,a;θ( )( ) 2 DQNの損失関数 ：experience replay memory ：ターゲット・ネットワーク D θ−
- 6. 強化学習の基本③ actor-critic法のシステム Value Function Policy Critic Environment Sutton, Berto. “Reinforcement Learning –an introduction.” 1998. state reward Actor TD error action
- 7. DQN(NIPs 2013)のしくみ Nair, et. al “Massively parallel methods for deep reinforcement learning.” In ICML Deep learning Workshop. 2015.
- 8. DQN(nature 2015)のしくみ Nair, et. al “Massively parallel methods for deep reinforcement learning.” In ICML Deep learning Workshop. 2015.
- 9. DistBeliefのしくみ J. Dean, et al “Large Scale Distributed Deep Networks.” NIPS. 2012. ここでcomputer間 のやりとり
- 10. Downpour SGDのしくみ J. Dean, et al “Large Scale Distributed Deep Networks.” NIPS. 2012. 最新のparameterを 保持するmaster replicaから要求があったら、 その時点の最新parameterを 返す → replicaによってある 時点で使ってるparameterが 違う → asyncronous SGDの計算が終 わったら勾配を返 す DistBeliefで勾配計 算 ミニバッチをreplicaご とに分割
- 11. Sandblaster L-BFGSのしくみ J. Dean, et al “Large Scale Distributed Deep Networks.” NIPS. 2012.
- 12. Gorilaのしくみ A. Nair, et al “Massively parallel methods for deep reinforcement learning.” In ICML Deep learning Workshop. 2015.
- 13. Gorilaのしくみ ver.1 共有のreplay memoryを使用 Environment Q Network Shard 1 Shard 2 Shard K Q Network Target Q Network DQN Loss Parameter Server Environment Q Network Q Network Target Q Network DQN Loss ・ ・ ・ ActorのcomputerとLearnerの computer１つずつで１セットとする Actor Learner 全部でNセット replay memoryは１ つを共有する Replay Memory
- 14. Gorilaのしくみ ver.2（bundled mode） 個別のreplay memoryを使用 Environment Q Network Shard 1 Shard 2 Shard K Q Network Target Q Network DQN Loss Replay Memory Parameter Server Environment Q Network Q Network Target Q Network DQN Loss Replay Memory ・ ・ ・ ActorのcomputerとLearnerの computer１つずつで１セットとする Actor Learner 全部でNセット replay memoryはそれぞれ のcomputerに配置
- 15. Gorila（bundled mode）から asynchronousなDQNへの変更点① Environment Q Network Shard 1 Shard 2 Shard K Q Network Target Q Network DQN Loss Replay Memory Parameter Server Environment Q Network Q Network Target Q Network DQN Loss Replay Memory ・ ・ ・ CPU上の１つのスレッドに対応 Actor Learner replay memoryを廃止
- 16. Gorila（bundled mode）から asynchronousなDQNへの変更点② Environment Q Network Shard 1 Shard 2 Shard K Q Network Target Q Network DQN Loss Parameter Server Environment Q Network Q Network Target Q Network DQN Loss ・ ・ ・ Actor Learner 代わりに勾配を溜め込む gradients gradients
- 17. Gorila（bundled mode）から asynchronousなDQNへの変更点③ Environment Q Network Shard 1 Shard 2 Shard K Q Network Target Q Network DQN Loss Parameter Server for Q-Network Environment Q Network Q Network Target Q Network DQN Loss ・ ・ ・ Actor Learner gradients gradients Shard 1 Shard 2 Shard K Parameter Server for Target Q-Network Target Q-Network用のserverを作る
- 18. Shard 1 Shard 2 Shard K Parameter Server for Q-Network Shard 1 Shard 2 Shard K Parameter Server for Target Q-Network AsynchronousなDQNの流れ① Environment Q Network Q Network Target Q Network DQN Loss Environment Q Network Q Network Target Q Network DQN Loss ・ ・ ・ Actor Learner θをコピー をコピー gradients gradients θ−
- 19. Shard 1 Shard 2 Shard K Parameter Server for Q-Network Shard 1 Shard 2 Shard K Parameter Server for Target Q-Network AsynchronousなDQNの流れ② Environment Q Network Q Network Target Q Network DQN Loss gradients Environment Q Network Q Network Target Q Network DQN Loss ・ ・ ・ Actor Learner 状態 s で行動 a をとり、s’ や r を観測 gradients
- 20. Shard 1 Shard 2 Shard K Parameter Server for Q-Network Shard 1 Shard 2 Shard K Parameter Server for Target Q-Network AsynchronousなDQNの流れ③ Environment Q Network Q Network Target Q Network DQN Loss gradients Environment Q Network Q Network Target Q Network DQN Loss ・ ・ ・ Actor Learner gradients L θ( )= Es,a,r,s'≈D r +γ max a' Q s',a';θ− ( )−Q s,a;θ( )( ) 2 Lossを計算
- 21. Shard 1 Shard 2 Shard K Parameter Server for Q-Network Shard 1 Shard 2 Shard K Parameter Server for Target Q-Network AsynchronousなDQNの流れ④ Environment Q Network Q Network Target Q Network DQN Loss gradients Environment Q Network Q Network Target Q Network DQN Loss ・ ・ ・ Actor Learner gradients 勾配を溜め込む dθ ← dθ + ∂L θ( ) ∂θ
- 22. Shard 1 Shard 2 Shard K Parameter Server for Q-Network Shard 1 Shard 2 Shard K Parameter Server for Target Q-Network AsynchronousなDQNの流れ⑤ Environment Q Network Q Network Target Q Network DQN Loss gradients Environment Q Network Q Network Target Q Network DQN Loss ・ ・ ・ Actor Learner gradients 定期的に勾配の積算値 を送り学習する dθ
- 23. A3Cのしくみ Environment Network Network gradients Environment ・ ・ ・ gradients Actor Critic Shard 1 Shard 2 Shard K Parameter Server for Network θ θv ' Network Shard 1 Shard 2 Shard K Parameter Server for Network θv gradients dθ dθv θ ' θ '
- 24. A3Cの流れ① Environment Network Network gradients Environment ・ ・ ・ gradients Actor Critic Shard 1 Shard 2 Shard K Parameter Server for Network θ θv ' Network Shard 1 Shard 2 Shard K Parameter Server for Network θv gradients dθ dθv θをコピー をコピー θv θ ' θ '
- 25. A3Cの流れ② Environment Network Network gradients Environment ・ ・ ・ gradients Actor Critic Shard 1 Shard 2 Shard K Parameter Server for Network θ θv ' Network Shard 1 Shard 2 Shard K Parameter Server for Network θv gradients dθ dθv ステップ間、状態 で方策 に従い行動 をとる。 tmax π at | st;θ '( ) at st V st,θv '( ) を計算する θ ' θ '
- 26. A3Cの流れ③ Environment Network Network gradients Environment ・ ・ ・ gradients Actor Critic Shard 1 Shard 2 Shard K Parameter Server for Network θ θv ' Network Shard 1 Shard 2 Shard K Parameter Server for Network θv gradients dθ dθv θ ' θ ' とそれぞれの勾 配を計算する R = γi−1 rt−i i=1 tmax ∑ +V st,θv '( ) dθ = ∇θ ' logπ ai | si;θ '( ) R −V si;θv '( )( ) dθv = dθv + ∂ R −V si;θv '( )( ) 2 ∂θv '
- 27. A3Cの流れ④ Environment Network Network gradients Environment ・ ・ ・ gradients Actor Critic Shard 1 Shard 2 Shard K Parameter Server for Network θ θv ' Network Shard 1 Shard 2 Shard K Parameter Server for Network θv gradients dθ dθv θ ' θ ' それぞれの勾配を溜め込む
- 28. A3Cの流れ④ Environment Network Network gradients Environment ・ ・ ・ gradients Actor Critic Shard 1 Shard 2 Shard K Parameter Server for Network θ θv ' Network Shard 1 Shard 2 Shard K Parameter Server for Network θv gradients dθ dθv θ ' θ ' ごとに勾配の積算値 を送り学習 dθtmax ごとに勾配の積算値 を送り学習 tmax dθv