RL Theory

[eleurent]

RL Theory is not properly represented. A new section should be added, with at least:

• Tabular setting
  • With a generative model
    • QVI
  • Without
    • UCRL2
    • UCBVI
  • Episodic
  • Q-learning+UCB
• Extensions to compact state-action spaces
• Extension to Kernels
• Performance measures: PAC, simple regret, cumulative regret, etc.
• RL with compatible function approximation

Is there a difference between generative models (sample any transition) and simulators (simulate trajectories from current states only)?