The Natural Language of Actions Guy Tennenholtz , Shie Mannor ICML - - PowerPoint PPT Presentation

the natural language of actions
SMART_READER_LITE
LIVE PREVIEW

The Natural Language of Actions Guy Tennenholtz , Shie Mannor ICML - - PowerPoint PPT Presentation

The Natural Language of Actions Guy Tennenholtz , Shie Mannor ICML 2019 Technion Institute of Technology What is Language? What is Language? Language is a purely human and non-instinctive method of communicating ideas, emotions, and desires


slide-1
SLIDE 1

The Natural Language of Actions

Guy Tennenholtz, Shie Mannor ICML 2019

Technion Institute of Technology

slide-2
SLIDE 2

What is Language?

slide-3
SLIDE 3

What is Language?

‘Language is a purely human and non-instinctive method of communicating ideas, emotions, and desires by means of voluntarily produced symbols.’ Edward Sapir, 1921

2

slide-4
SLIDE 4

What is Language?

3

slide-5
SLIDE 5

Contextual Semantics

Contextual Representation A word’s contextual representation is an abstract congnitive structure that emmulates from encounters with the word in various linguistic contexts. We learn new words based on contextual cues ∙ I saw a little yazuba sleeping behind the tree. ∙ One glass of feandra is enough to get you drunk.

4

slide-6
SLIDE 6

Meaning of Actions?

Characteristics ∙ Actions characterized by the company they keep ∙ Context is generated by an

  • ptimal agent

∙ Agent demonstrates acceptable behavior in the environment

5

slide-7
SLIDE 7

Act2Vec

slide-8
SLIDE 8

Act2Vec

7

slide-9
SLIDE 9

Skip Gram (used in Word2Vec)

8

slide-10
SLIDE 10

Skip Gram (used in Word2Vec)

9

slide-11
SLIDE 11

Act2Vec: Drawing

∙ 70,000 human drawn squares from the QuickDraw! dataset ∙ Strokes correspond to two

  • perations

10

slide-12
SLIDE 12

Act2Vec: Navigation

Act2Vec embedding trained in 2d environment on actions sequences of primitive actions: ∙ Move Forward 1 unit ∙ Turn Left 15 degrees ∙ Turn Right 15 degrees

11

slide-13
SLIDE 13

Act2Vec: Navigation

12

slide-14
SLIDE 14

Act2Vec in Reinforcement Learning

slide-15
SLIDE 15

Act2Vec in Reinforcement Learning

14

slide-16
SLIDE 16

Act2Vec for Q-Value Approximation

15

slide-17
SLIDE 17

Q-Embedding

1

1He, Ji, et al. ”Deep Reinforcement Learning with a Natural Language Action Space.” Proceedings of

the 54th Annual Meeting of the Association for Computational Linguistics (2016).

16

slide-18
SLIDE 18

Act2Vec for Exploration

17

slide-19
SLIDE 19

k-Exp

k-Exp ∙ Divide action embedding space into k clusters using a clustering algorithm (e.g., k-means) ∙ Sample a cluster uniformly ∙ Given a cluster, uniformly sample an action within it

18

slide-20
SLIDE 20

Act2Vec for Domain Transfer

19

slide-21
SLIDE 21

Navigation

Transfer to 3d environment

20

slide-22
SLIDE 22

The Semantics of Actions in Starcraft II

slide-23
SLIDE 23

StarCraft II

22

slide-24
SLIDE 24

StarCraft II

23

slide-25
SLIDE 25

StarCraft II

24

slide-26
SLIDE 26

Conclusion

slide-27
SLIDE 27

Conclusion

∙ Actions are thought of as symbols of a natural, expressive language ∙ Actions represent thoughts, beliefs, and strategies ∙ Their meaning is captured through their context ∙ Similarity between actions helps reduce cardinality of action space ∙ Prior knowledge is incorporated through action representations

26