Collaborative Evolutionary Reinforcement Learning Shauharda Khadka, - - PowerPoint PPT Presentation

collaborative evolutionary reinforcement learning
SMART_READER_LITE
LIVE PREVIEW

Collaborative Evolutionary Reinforcement Learning Shauharda Khadka, - - PowerPoint PPT Presentation

Collaborative Evolutionary Reinforcement Learning Shauharda Khadka, Somdeb Majumdar, Tarek Nassar, Zach Dwiel, Evren Tumer, Santiago Miret, Yinyin Liu, Kagan Tumer* Artificial Intelligence Products Group, Intel Corporation Oregon State


slide-1
SLIDE 1

Collaborative Evolutionary Reinforcement Learning

Shauharda Khadka, Somdeb Majumdar, Tarek Nassar, Zach Dwiel, Evren Tumer, Santiago Miret, Yinyin Liu, Kagan Tumer* Artificial Intelligence Products Group, Intel Corporation Oregon State University*

slide-2
SLIDE 2

A simple actor-critic policy gradient setup

slide-3
SLIDE 3

Learner

slide-4
SLIDE 4

What do we optimize exactly?

slide-5
SLIDE 5

Learner

slide-6
SLIDE 6

Portfolio of Learners (varying discount rates)

slide-7
SLIDE 7

Why varying discount rates?

slide-8
SLIDE 8

Why varying discount rates?

slide-9
SLIDE 9

Back to Portfolio of Learners

slide-10
SLIDE 10

Adding a Resource Manager

slide-11
SLIDE 11

Adding Neuroevolution

slide-12
SLIDE 12

12

Experiment: Humanoid

slide-13
SLIDE 13

13

Experiment: Humanoid

  • Solves Humanoid under 1 million samples
  • TD3 learners fail entirely
  • Neuroevolution ~62.5 million samples
slide-14
SLIDE 14