Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Deep reinforcement learning : Starcraft learning environment by Gema Parreño at Big Data Spain 2017

Deep reinforcement learning : Starcraft learning environment by Gema Parreño at Big Data Spain 2017

A theorical description of reinforcement learning principles and a deep dive into DeepMind Research environment .

https://www.bigdataspain.org/2017/talk/reinforced-learning-deepmind-starcraft-learning-environment

Big Data Spain 2017
November 16th - 17th Kinépolis Madrid

Cb6e6da05b5b943d2691ceefa3381cad?s=128

Big Data Spain

December 01, 2017
Tweet

Transcript

  1. None
  2. sTARcRAFT ii Deep Reinforced Learning Gema Parreño Piqueras @SoyGema

  3. Supervised Learning

  4. LEARNING by known data

  5. Unsupervised Learning

  6. LEARNING by unknown data

  7. Reinforce Learning

  8. Classify Machine Learning by learning method

  9. Reinforce Learning LEARNING WHILE INTERACTING WITH ENVIRONMENT

  10. States / Actions

  11. ENVIRONMENT

  12. Agent

  13. Policy

  14. States / Actions

  15. The ATARI CASE

  16. Environment description x 2600

  17. Policy Architecture

  18. Policy Architecture Input SPATIAL/ NON SPATIAL FEATURES

  19. Policy Architecture Output SPATIAL/ NON SPATIAL ACTION POLICY

  20. Results

  21. Conclusion Meta solution

  22. Challenges

  23. TRIAL AND ERROR

  24. EXPLORATION VS EXPLOTATION

  25. STARCRAFT 2

  26. What is sc2 and why?

  27. Learning Environmnet

  28. Policy Search

  29. Agent Architectures

  30. Mini games

  31. Mini-game definition

  32. Agent without training Agent trained

  33. None
  34. Sentry unit

  35. Designing environment

  36. The Maps 1st Iteration

  37. The Maps 2nd Iteration

  38. The Maps 3th Iteration

  39. Results Results

  40. Diferent type of Nets On random agent

  41. Diferent type of Nets On scripted agent

  42. The future

  43. RL out of Videogames

  44. The environment The environment

  45. The GOAL The GOAL

  46. The policy The Policy

  47. Thanks! Deep Reinforced Learning Gema Parreño Piqueras @SoyGema