Rise of unity_ml_7_22
- 1. Rise of the Reinforcement Learning Wonseok Jung
- 2. 螳 伎襴暑 - Baruch college (Data Science Major) ConnexionAI Researcher CTRL (Contest in RL) 襴 DeepLearningCollege 螳 郁規 Project : Object Detection, Chatbot, Reinforcement Learning Github: https://github.com/wonseokjung Facebook: https://www.facebook.com/ws.jung.798 Blog: https://wonseokjung.github.io/
- 3. 覈谿 1. Create Environments 2. Multi-Agent Environment 3.Adversarial self-play 4. Imitation Learning 5. Curriculum Learning
- 5. - 螳讌 蟆曙 蠏 蟆曙 襷 螳 螻襴讀 覲伎. - 螻殊 螳讌 伎螳 覦. OpenAI-gym DQN Supermario DDQN(tuned) Sonic Rainbow DQN(tuned) OpenSim DDPG
- 6. QUESTIONS 讌覓 : 譯殊伎 蟆曙 願 蟠蠍 覓語襯 蠍 蟆曙 襷 蟆 螳ロ蟾? 旧螳 譴殊 覦覯 蟾? Issues : 1. 旧螳 覓企覓 る蟇碁Π. I 8, 1080蠍一 : OpenAI GYM : 豕 5覿 ~ 殊殊 伎 SuperMario Level 1 : 6 Sonic : OpenAI 螻 覯 : 7螳 Prosthetics : 1 伎 2. 螳旧 旧 . 螻牛 蟆暑 蟆 螳ロ.
- 7. UNITY ML-AGENTS - Unity襯 螳語 螳 蟆曙 襷れ . - Machine Learning Agents 蠍磯レ朱 旧 覲企 螻殊願 蟆 螳ロ.
- 9. MULTI-AGENTS? -Intelligent human agents() るジ agents 覲企ゼ 螻旧. -覲企ゼ 螻旧覃 cooperation() 蟇磯 Independent(襴曙朱) 蟆 覈襯 燕.
- 10. MULTI-AGENTS - 豐 Agents螳 . - 覦襯 penalty襯 覦朱 , 碁 覦襯 Reward襯 覦 . - 螳 Agent 襴暑 Brain 螳讌螻 朱 襴曙朱 action .
- 11. TRAINING USING IMITATION LEARNING - Agents Independent蟆 碁Banana襯 谿剰鍵 螳讌 action 覃 覦一企.
- 13. ADVERSARIAL LEARNING? - 螻牛給 覈襯 燕蠍 覲企ゼ 螻旧覃 Cooperation 讌襷, 覲旧, 豢蟲, 蟲, 炎骸 螳 麹螳 ろ 蟆曙磯 .
- 14. ADVERSARIAL LEARNING - 螻牛給 覈襯 燕蠍 覲企ゼ 螻旧覃 Cooperation 讌襷, 覲旧, 豢蟲, 蟲, 炎骸 螳 麹螳 ろ 蟆曙磯 . - 螻 ｋ striker 螻 襷 Goalkeeper襦 蟲焔 . Striker Goalkeeper Striker Goalkeeper Object
- 15. ADVERSARIAL LEARNING - Striker Goalkeeper ｸ 螻旧 ｊ 襷蠍 Cooperation覃 . Striker Goalkeeper GoalKeeper VS StrikerVS Coop Coop
- 17. IMITATION LEARNING? - 螻 覓殊 企 覈覓殊 覲願 蠏碁れ behavior 覲願 覦一企. - 蟆 覈覦覃 覦一磯 覦覯 Imitation Learning企手 .
- 18. TRAINING USING IMITATION LEARNING Gravity Agent1 Agent2 Gravity Ball Ball Initialization - 螻旧 譴レ 伎覃, 螳 Agent 螻旧 覦 覦 Agent 朱 蟆狩. - 覦 豺 Agent 螻旧 覲願 れ 蠍企.
- 19. TRAINING USING IMITATION LEARNING Agent1 Agent2 Strat Training Action3 Action1Action2 Action3 Action1 Action3 - Agent 螳讌 action 覃 襷 Reward襯 覦 action - 企 覦 給 螳 襷 .
- 20. TRAINING WITHOUT IMITATION LEARNING - 襦覺 蟇穴 郁 襷蟆螻 螳 企れ 蟆曙 旧螳 襷れ 蠍碁 一る曙 覓語螳 覦.
- 21. TRAINING USING IMITATION LEARNING Imitation Learning Teacher Student - 觜襯願 螻殊朱 覦一語 蟆 覓瑚 覲願 覦一磯 Imitation Learning . - Student Teacher 覲願 覦一 - Teacher(Player) action 覃 student螳 觜襴 覦一語 蟆 譴.
- 23. CURRICULUM LEARNING - 企 覦襦 襴 蟆 螳ロ讌 . - れ螻, 蠍郁, 螻, 蟇穴, 襴 蟆豌 螻覲襦 牛. - 企 覦覯 螳旧 Curriculum Learning企手 .
- 24. CURRICULUM LEARNING - Agent 螳 task覿 覦一郁鍵 螻覲襦 旧 . - 覯 企れ task襯 牛蠍 り鍵 覓語 螻襯 旧 .
- 25. ENVIRONMENTS Agent Goal Wall - Agent Goal 谿蠍 action . - Wall れ螳讌 企 覦. Action1Action2 Action1
- 26. TRAINING USING IMITATION LEARNING - agent small wall螻 large wall 狩蟇磯 一企 覈襦 螳蠍 牛.
- 27. SUMMARY 1. Create Environments 2. Multi-Agent Environment 3.Adversarial self-play 4. Imitation Learning 5. Curriculum Learning