Policy Gradient Methods

  • learns state to action mapping directly which is often more simple
  • no model of environment dynamics needed
  • allows continuous action space
  • allows stochastic policy which can be a crucial advantage compared to deterministic policies
  • Actor Critic RL Methods

Reinforcement Learning, Policy Gradients and Actor Critics Hado Van Hasselt - Reinforcement Learning, Policy Gradients and Actor Critics (2018)