The document discusses two recent papers on off-policy meta-reinforcement learning: 1) "Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables" which introduces PEARL, an off-policy method for meta RL using context variables to enable efficient adaptation. 2) "Guided Meta-Policy Search" which uses a two-level approach of task learning and meta-learning, where task learning trains policies via RL and meta-learning trains a meta-objective via imitation. Both papers aim to enable efficient off-policy adaptation in meta RL.

1. 1
off-policy
Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables
Guided Meta-Policy Search
Presenter:Tatsuya Matsushima @__tmats__ , Matsuo Lab

2. • off-policy arXiv
• Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables
[Rakelly+ 2019] (2019/3/19)
• Guided Meta-Policy Search [Mendonca+ 2019] (2019/4/1)
• MAML meta-training
off-policy
2

3. 3

4. (meta learning)
• : Wiki http://ibisforest.org/index.php?%E3%83%A1%E3%82%BF%E5%AD%A6%E7%BF%92
• [DL ]Meta-Learning Probabilistic Inference for Prediction ( )
• https://www.slideshare.net/DeepLearningJP2016/dlmetalearning-probabilistic-inference-for-
prediction-126167192
4

5. MAML
MAML (Model Agnostic Meta-Learning) [Finn+ 2017]
•
• adapt
• MAML
• 2
• 1 [Nichol+2018]
• meta-test
5
min
θ ∑
𝒯
ℒ (θ − α∇θℒ (θ, 𝒟tr
𝒯), 𝒟val
𝒯 ) = min
θ ∑
𝒯
ℒ (ϕ 𝒯, 𝒟val
𝒯 )
θ ϕ 𝒯
ϕ 𝒯test
= θ − α∇θℒ (θ, 𝒟tr
𝒯test)

6. MAML
• loss loss( )
• MAML model-based [Nagabandi+ 2018] [Gupta+ 2018]
• [DL ]Meta Reinforcement Learning ( )
• https://www.slideshare.net/DeepLearningJP2016/dl-130067084
6
ℒRL (ϕ, 𝒟 𝒯i) = −
1
𝒟 𝒯i
∑
st,at∈𝒟
ri (st, at)
= − 𝔼st,at∼πϕ,q 𝒯i [
1
H
H
∑
t=1
ri (st, at)
]

7. ( ) On-policy v.s. Off-policy
On-policy ( )
• ( )
•
• ) ε-greedy
Off-policy ( )
•
•
※ MAML train
test (= off-policy )
7

8. Efficient Off-Policy Meta-Reinforcement
Learning via Probabilistic Context Variables
8

9. Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context
Variables
• https://arxiv.org/abs/1903.08254 (Submitted on 19 Mar 2019)
• Kate Rakelly, Aurick Zhou, Deirdre Quillen, Chelsea Finn, Sergey Levine
• UC Berkeley (BAIR)
• Deep RL ”UC Berkeley”
•
• https://github.com/katerakelly/oyster
• (BAIR )PyTorch rlkit
9

10. TL; DR
• meta learning off-policy (PEARL)
• (context)
• permutation invariant
• 20-100
10

11. ( MAML )
• meta-training adaptation on-policy
• MAML meta-train meta-test off-policy
• adapt
•
11

12. 12

13. • off-policy RL (soft actor-critic, SAC [Haarnoja+ 2018])
context (PEARL)
• Meta-training adapt
• meta-train policy
context
• meta-test context policy
adapt
• policy off-policy meta-train meta-
test on-policy
13

14. MDP
•
•
•
• :
• :
• 1
•
•
14
p(𝒯)
𝒯 𝒯 = {p (s0), p (st+1 |st, at), r (st, at)}
𝒯 c 𝒯
n = (sn, an, rn, s′n)
c = c 𝒯
1:N
p(𝒯)

15. context
• adapt
•
• (Inference network)
•
• prior Gaussian
• meta-train meta-test
15
z
z
qϕ(z|c)
𝔼 𝒯
[
𝔼z∼qϕ(z|c 𝒯
) [
R(𝒯, z) + βDKL (qϕ (z|c 𝒯
) ∥p(z))]]
p(z)
qϕ(z|c) ϕ zz

16. context
• MDP
• permutation invariant
• Inference network
• Gaussian
16
{si, ai, s′i, ri}
qϕ (z|c1:N) ∝ ΠN
n=1Ψϕ (z|cn)
Ψϕ (z|cn) = 𝒩 (fμ
ϕ (cn), fσ
ϕ (cn))

17. off-policy
• policy
• actor ciritic
•
• on-policy
on-policy test
17
qϕ(z|c)
ℬ
𝒮c

18. off-policy
• Soft Actor-Critic (SAC) [Haarnoja+ 2018] context
• SAC maxEntRL( ) off-policy actor-critic
• actor critic reparameterization trick
• critic loss:
• actor loss:
18
ℒcritic = 𝔼(s, a, r, s′
) ∼ ℬ
z ∼ qϕ(z|c)
[Qθ(s, a, z) − (r + V (s′, z))]
2
z
ℒactor = 𝔼s∼ℬ,a∼πθ
DKL
(
πθ(a|s, z)∥
exp (Qθ(s, a, z))
𝒵θ(s) )

19. 19

20. • MuJoCo 6
• Half-Cheetah, Humanoid, Ant, Walker (Half-Cheetah Ant 2 )
•
• adapt
• 20-100
• : meta-training
• :
20

21. • on-policy (MAESN[Gupta+ 2018])
• sparse navigation
• meta-test
•
• context
• MAESN
21

22. Ablation Study
•
• Half-Cheetah-Vel
• RNN
• RNN-tran: de-correlated
• RNN-traj:
• permutation invariant
22

23. Ablation Study
•
• Half-Cheetah-Vel
•
• off-policy: off-policy( )
• off-policy RL-batch: policy
•
(PEARL)
23

24. Ablation Study
• context
• sparse navigation
• context
•
24

25. 25

26. • off-policy (PEARL)
• context policy context
off-policy
• meta-training
26

27. Guided Meta-Policy Search
27

28. Guided Meta-Policy Search
• https://arxiv.org/abs/1904.00956 (Submitted on 1 Apr 2019)
• Russell Mendonca, Abhishek Gupta, Rosen Kralev, Pieter Abbeel, Sergey Levine, Chelsea
Finn
• UC Berkeley (BAIR)
• …
•
• https://github.com/RussellM2020/GMPS
• Website
• https://sites.google.com/berkeley.edu/guided-metapolicy-search
28

29. TL; DR
• meta learning off-policy (GMPS)
• meta-train RL
• meta-train meta-objective( ) imitation learning (behaviour cloning)
• meta-training task learning meta-learning 2
29

30. ( MAML )
• meta-training adaptation on-policy
• [Rakelly+ 2019]
• meta-training meta-test
30

31. 31

32. • meta-train meta-objective( ) (behaviour cloning)
• meta-training 2
• task learning: meta-training policy
• policy meta-test expert
• meta-learning: policy meta-level supervised
32

33. [Rakelly+ 2019]
•
•
•
33
p(𝒯)
𝒯 𝒯 = {p (s0), p (st+1 |st, at), r (st, at)}
p(𝒯)

34. task learning
• meta-training
/ policy
•
meta-learning
• MAML
• adapt
• MAML
• (behaviour cloning)
34
𝒯i
{π*i
}
ℒRL (ϕi, 𝒟i)
ϕi 𝒯i
ℒBC (ϕi, 𝒟i) ≜ −
∑
(st,at)∈𝒟
log πϕ (at |st)

35. meta-learning
• meta-training
• policy
meta-objective
•
behaviour cloning compounding error
35
𝒯i
π*i
D*i
min
θ ∑
𝒯i
∑
𝒟val
i ∼𝒟*i
𝔼 𝒟tr
i ∼πθ [
ℒBC (θ − α∇θℒRL (θ, 𝒟tr
i ), 𝒟val
i )]
θ
𝒯i
ϕi
D*i

36. • meta-learning task learning meta-learning
• policy
•
• meta-training
• ) reward shaping
• MAML
36

37. policy
• policy
contextual policy
• ( ID )
• meta-training
• meta-test meta-training
• soft actor-critic(SAC) [Haarnoja+ 2018]
37
πθ (at |st, ω)
ω

38. • Behaviour cloning meta-objective
•
•
•
• Behaviour cloning
38
θ
ϕi
πθ
ϕi = θ + α𝔼τ∼πθ
[
πθ(τ)
πθinit
(τ)
∇θlog πθ(τ)Ai(τ)
]
Ai
θ ← θ − β∇θℒBC (ϕi, 𝒟val
i )

39. 39

40. •
• Pushing (full state)
•
•
• Pushing (vision)
•
• Door opening
•
•
• (Ant)
•
https://sites.google.com/berkeley.edu/guided-metapolicy-search 40

41. •
• meta-training task context( )
• SAC
• : meta-training :
41

42. •
• Door Opening Ant
•
• pushing
•
42

43. 43

44. • off-policy (GMPS)
• meta-training task learning meta-learning 2
(behaviour cloning)
• meta-training
44

45. 45

46. • 2
• one-step update adapt (BAIR )
• ) MAML[Finn+ 2017]
• adapt (DeepMind )
• ) Neural Processes[Garnelo+ 2018], GQN[Eslami+ 2018]
•
•
• [DL ]Meta-Learning Probabilistic Inference for Prediction
• https://www.slideshare.net/DeepLearningJP2016/dlmetalearning-probabilistic-inference-for-
prediction-126167192
• pro-con
46

47. Appendix
47

48. References
[Eslami+ 2018] Eslami, S. M. Ali, Danilo Jimenez Rezende, Frédéric Besse, Fabio Viola, Ari S. Morcos, Marta Garnelo, Avraham Ruderman,
Andrei A. Rusu, Ivo Danihelka, Karol Gregor, David P. Reichert, Lars Buesing, Theophane Weber, Oriol Vinyals, Dan Rosenbaum, Neil C.
Rabinowitz, Helen King, Chloe Hillier, Matthew M Botvinick, Daan Wierstra, Koray Kavukcuoglu and Demis Hassabis. “Neural scene
representation and rendering.” Science 360 (2018): 1204-1210. http://science.sciencemag.org/content/360/6394/1204
{Finn+ 2017] Chelsea Finn, Pieter Abbeel and Sergey Levine. “Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks,”
Proceedings of the 34th International Conference on Machine Learning, PMLR 70:1126-1135, 2017. http://proceedings.mlr.press/v70/
finn17a.html
[Garnelo+ 2018] Marta Garnelo, Jonathan Schwarz, Dan Rosenbaum, Fabio Viola and Danilo J. Rezende, S.M. Ali Eslami and Yee Whye
Teh. “Neural Processes”. https://arxiv.org/abs/1807.01622.
[Gupta+ 2018] Abhishek Gupta, Russell Mendonca, YuXuan Liu, Pieter Abbeel and Sergey Levine. ”Meta-Reinforcement Learning of
Structured Exploration Strategies”. In Advances in Neural Information Processing Systems, 2018. https://nips.cc/Conferences/2018/
Schedule?showEvent=12658
[Haarnoja+ 2018] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel and Sergey Levine. “Soft Actor-Critic: Off-Policy Maximum Entropy Deep
Reinforcement Learning with a Stochastic Actor”. Proceedings of the 35th International Conference on Machine Learning, PMLR
80:1861-1870, 2018. http://proceedings.mlr.press/v80/haarnoja18b.html
[Mendonca+ 2019] Russell Mendonca, Abhishek Gupta, Rosen Kralev, Pieter Abbeel, Sergey Levine and Chelsea Finn. “Guided Meta-
Policy Search”. https://arxiv.org/abs/1904.00956
[Nagabandi+ 2018] Anusha Nagabandi, Ignasi Clavera, Simin Liu, Ronald S. Fearing, Pieter Abbeel, Sergey Levine and Chelsea Finn.
“Learning to Adapt in Dynamic, Real-World Environments Through Meta-Reinforcement Learning”. https://arxiv.org/abs/1803.11347
[Nichol+2018] Alex Nichol, Joshua Achiam and John Schulman. “On First-Order Meta-Learning Algorithms”. https://arxiv.org/abs/1803.02999
[Rakelly+ 2019] Kate Rakelly, Aurick Zhou, Deirdre Quillen, Chelsea Finn ands Sergey Levine. “Efficient Off-Policy Meta-Reinforcement
Learning via Probabilistic Context Variables”. https://arxiv.org/abs/1903.08254
48