hindsight credit assignment
play

Hindsight Credit Assignment Anna Harutyunyan , Will Dabney, Thomas - PowerPoint PPT Presentation

Hindsight Credit Assignment Anna Harutyunyan , Will Dabney, Thomas Mesnard, Mo Azar, Bilal Piot, Nicolas Heess, Hado van Hasselt, Greg Wayne, Satinder Singh, Doina Precup, Remi Munos NeurIPS 2019 How did past actions infmuence future


  1. Hindsight Credit Assignment Anna Harutyunyan , Will Dabney, Thomas Mesnard, Mo Azar, Bilal Piot, Nicolas Heess, Hado van Hasselt, Greg Wayne, Satinder Singh, Doina Precup, Remi Munos NeurIPS 2019

  2. How did past actions infmuence future outcomes?

  3. RL relies on MDP structure, and takes time as main proxy for credit relevance

  4. RL relies on MDP structure, and takes time as main proxy for credit relevance

  5. RL relies on MDP structure, and takes time as main proxy for credit relevance Time

  6. RL relies on MDP structure, and takes time as main proxy for credit relevance Credit Time

  7. RL relies on MDP structure, and takes time as main proxy for credit relevance Credit Meeting at 4pm Time

  8. RL relies on MDP structure, and takes time as main proxy for credit relevance Credit Lunch Time

  9. RL relies on MDP structure, and takes time as main proxy for credit relevance Credit Umbrella Time

  10. RL relies on MDP structure, and takes time as main proxy for credit relevance Credit Time Credit

  11. Instead of only relying on MDP assumptions, let’s learn credit relevance explicitly!

  12. x

  13. x past future action outcome

  14. x past future action outcome

  15. How did past actions infmuence future outcomes? x past future action outcome

  16. How did past actions infmuence future outcomes? x y past future action outcome

  17. How did past actions infmuence future outcomes? x y past future action outcome

  18. How did past actions infmuence future outcomes? x y past future action outcome

  19. How did past actions infmuence future outcomes? x y past future action outcome

  20. How did past actions infmuence future outcomes? z x y past future action outcome

  21. How did past actions infmuence future outcomes? z x y past future action outcome

  22. How did past actions infmuence future outcomes? z x y past future action outcome

  23. How did past actions infmuence future outcomes? z x y past future action outcome State

  24. How did past actions infmuence future outcomes? z x y past future action outcome State Return

  25. Hindsight Credit Assignment How relevant was a to get to a state X k ?

  26. Hindsight Credit Assignment How relevant was a to get to a state X k ? How relevant was a to achieve the return Z ?

  27. Hindsight Credit Assignment How relevant was a to get to a state X k ? How relevant was a to achieve the return Z ?

  28. Hindsight Credit Assignment HCA Algorithms: Learn the hindsight distribution P, and use it to better estimate value functions or policy gradients How relevant was a to get to a state X k ? How relevant was a to achieve the return Z ?

  29. Experiments

  30. Experiments

  31. Thank you for your attention! Poster #204 :)

Recommend


More recommend