Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle...
Transcript of Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle...
![Page 1: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/1.jpg)
![Page 2: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/2.jpg)
Artificial Intelligence & Deep Learning
![Page 3: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/3.jpg)
Artificial Intelligence
3
— Classification
— Regression
— Natural Language Processing
— Object Detection
— Generative Model
— Reinforcement Learning
![Page 4: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/4.jpg)
4
Deep Learning
![Page 5: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/5.jpg)
Processing unit of Brain
Deep Learning
5
Processing unit of Neural Network
Input Output
![Page 6: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/6.jpg)
Deep Learning
6
~ 2010
![Page 7: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/7.jpg)
Deep Learning
7
Algorithms GPU Big Data
![Page 8: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/8.jpg)
Deep Learning
8
2010~
![Page 9: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/9.jpg)
Classification
9
![Page 10: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/10.jpg)
Regression
10
![Page 11: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/11.jpg)
Natural Language Processing
11
— OpenAI’s GPT-2
![Page 12: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/12.jpg)
Object Detection
12
![Page 13: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/13.jpg)
Generative Model
13
![Page 14: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/14.jpg)
Generative Model
14
![Page 15: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/15.jpg)
Generative Model
15
![Page 16: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/16.jpg)
Reinforcement Learning
16
![Page 17: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/17.jpg)
Deep Learning in Games
![Page 18: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/18.jpg)
Generative Adversarial Networks (GAN)
Deep Learning in Games
18
Reinforcement Learning (RL)
VS
![Page 19: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/19.jpg)
GAN
19
Real?
Fake?
Real
Fake
![Page 20: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/20.jpg)
GAN
20
Real
Fake
Real?
Fake?
Generator
Discriminator
![Page 21: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/21.jpg)
GAN
21
Training
![Page 22: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/22.jpg)
GAN
22
Generator
![Page 23: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/23.jpg)
Game Level Generation Using GAN
23
![Page 24: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/24.jpg)
Game Level Generation Using GAN
24
![Page 25: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/25.jpg)
Game Level Generation Using GAN
25
![Page 26: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/26.jpg)
Design Using GAN
26
Generator
Condition
![Page 27: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/27.jpg)
Design Using GAN
27
Generator
Sword
![Page 28: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/28.jpg)
Reinforcement Learning
28
Reward
+ Reward - Reward
![Page 29: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/29.jpg)
Reinforcement Learning
29
Reward
+ Reward - Reward
![Page 30: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/30.jpg)
Reinforcement Learning
30
Agent Environment
State, Reward
Action
![Page 31: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/31.jpg)
Reinforcement Learning
31
GridWorld Starcraft2
![Page 32: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/32.jpg)
Reinforcement Learning
32
![Page 33: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/33.jpg)
Reinforcement Learning
33
![Page 34: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/34.jpg)
Reinforcement Learning
34
![Page 35: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/35.jpg)
Reinforcement Learning
35
![Page 36: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/36.jpg)
Reinforcement Learning
36
![Page 37: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/37.jpg)
Reinforcement Learning
37
https://github.com/reinforcement-learning-kr/alpha_omok
![Page 38: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/38.jpg)
Reinforcement Learning
38
![Page 39: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/39.jpg)
Reinforcement Learning
39
![Page 40: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/40.jpg)
Reinforcement Learning
40
![Page 41: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/41.jpg)
Reinforcement Learning
41
— Multi-agents RL
— Meta RL
— Exploration
— Curiosity
— Noise in parameter
— Model-based RL
— Sim2Real
![Page 42: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/42.jpg)
Reinforcement Learning
42
https://www.facebook.com/groups/ReinforcementLearningKR/
![Page 43: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/43.jpg)
AI in Unity
![Page 44: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/44.jpg)
AI in Unity
44
Challenges Machine Learning Agents
![Page 45: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/45.jpg)
ML-agents Challenge
45
![Page 46: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/46.jpg)
ML-agents Challenge
46
![Page 47: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/47.jpg)
ML-agents Challenge
47
![Page 48: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/48.jpg)
Obstacle Tower Challenge
48
![Page 49: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/49.jpg)
Obstacle Tower Challenge
49
— Montezuma’s Revenge
— Challenges— Sparse reward
— Hard exploration
— Requires planning
— Multi Task
Hard to Solve!!
![Page 50: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/50.jpg)
Obstacle Tower Challenge
50
— Montezuma’s Revenge is Solved!
— Demonstration— Aytar et al. 2018
— Curiosity
– Pathak er al. 2017
– Burda et al. 2018
— Go-Explore
– Ecoffet et al. 2018
Go-Explore: a New Approach for Hard-Exploration Problems (Ecoffet et al.)
![Page 51: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/51.jpg)
Obstacle Tower Challenge
51
![Page 52: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/52.jpg)
Obstacle Tower Challenge
52
![Page 53: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/53.jpg)
Obstacle Tower Challenge
53
— 3D Visual Observation
— Complex floor layout
— Generalization
– Floor, Room, Wall
– Every 10 floors
— Multi Task
– Key
– Sokoban
– Pit
![Page 54: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/54.jpg)
Obstacle Tower Challenge
54
— ModuLabs CTRL Team— Kyushik Min
— Jay Jung
— Suhyuk Park
— Hyojeong Jeon
— Round 1 is finished! ☺
— Curiosity based algorithm
— Average 6~7th floor
![Page 55: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/55.jpg)
Obstacle Tower Challenge
55
![Page 56: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/56.jpg)
Unity ML-agents
56
![Page 57: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/57.jpg)
Unity ML-agents
57
— Reinforcement Learning
Agent Environment
State, Reward
Action
![Page 58: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/58.jpg)
Unity ML-agents
58
— Deep Reinforcement Learning
Agent Environment
?
State, Reward
Action
![Page 59: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/59.jpg)
Unity ML-agents
59
— Deep Reinforcement Learning
Agent Environment
?
State, Reward
Action
![Page 60: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/60.jpg)
Unity ML-agents
60
— Deep Reinforcement Learning
Agent Environment
State, Reward
Action
![Page 61: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/61.jpg)
Unity ML-agents
61
![Page 62: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/62.jpg)
Academy
- Managing brains
- Configuration setting
Brain
- Observation setting
- Action setting
Agent
- Script for Agent
- Control Setting
- Reward, done setting
Unity ML-agents
62
![Page 63: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/63.jpg)
Unity ML-agents
63
VS: Single Agent
: Multi-Agent
: Adversarial Agents
: Imitation Learning
![Page 64: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/64.jpg)
Unity ML-agents
64
: Training
: Heuristic
: PlayerBrain
: External
: Internal
![Page 65: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/65.jpg)
Unity ML-agents
65
Agent1
Brain1 (Heuristic)
Agent2
Brain2 (Internal)
Agent4
Brain3 (External)
Agent3 Agent5
Academy
![Page 66: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/66.jpg)
Deep Learning
Unity ML-agents
66
Reinforcement Learning Unity
![Page 67: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/67.jpg)
Unity ML-agents Tutorial
67
RL Korea Unity ML-agents Tutorial Team
https://github.com/reinforcement-learning-kr/Unity_ML_Agents
![Page 68: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/68.jpg)
Sokoban
- Discrete Action
- Deep Q-Network (DQN)
Unity ML-agents Tutorial
68
Drone
- Continuous Action
- Deep Deterministic Poli
cy Gradient (DDPG)
Pong
- Adversarial Environment
- Discrete Action
- DQN
![Page 69: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/69.jpg)
Sokoban (Curriculum)
- Curriculum Learning
- Discrete Action
- DQN
Unity ML-agents Tutorial
69
Dodge
- Imitation Learning
- Discrete Action
- Behavioral Cloning
TwoLeg Walker
- Multi-agents
- Continuous Action
- MADDPG
![Page 70: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/70.jpg)
Unity ML-agents Tutorial
70
![Page 71: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/71.jpg)
Unity ML-agents Tutorial
71
![Page 72: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/72.jpg)
Unity ML-agents Tutorial
72
![Page 73: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/73.jpg)
Unity ML-agents Tutorial
73
![Page 74: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/74.jpg)
Unity ML-agents Tutorial
74
![Page 75: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/75.jpg)
Unity ML-agents Tutorial
75
![Page 76: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/76.jpg)
Unity ML-agents Tutorial
76
![Page 77: Artificial Intelligence - Unite Seouluniteseoul.com/2019/PDF/D2T5S4.pdf · 2020. 2. 6. · Obstacle Tower Challenge 50 — Montezuma’s Revenge is Solved! — Demonstration — Aytar](https://reader035.fdocuments.net/reader035/viewer/2022071607/61442136aa0cd638b460a80a/html5/thumbnails/77.jpg)
Thank you
77
https://www.facebook.com/groups/ReinforcementLearningKR/