nickjalbert/studious-winner

studious-winner

• virtualenv -p which python3 studious-winner

How a brain trains a policy

It's a spectrum between:

• Hardcoded by evolution
• Imitation learning
• Prior learnings for the world
• Direct Explorations

Concrete next step:

• Agent is in world

• Agent explores but learns nothing

• Agent explores and accumulates priors about the world

• Build out apple world to be an explorable space

  • layer of indirection between reward and action

• Build out agent ability to explore

• Once the world is explorable, you learn things:

  • to satisfy drives
  • just in case later tasks require early learnings

Mapping between agent <-> MDPs and policies.

Reward for satisfying hunger and explore drives (internal)

Andy parting ideas ->

1. Bayes -- building model, updating model
2. ExploreDrive - two different ways (explore broadly vs explore to satisfy drives)
3. sergey lecture on meta-RL