Actor-critic is implicitly biased towards high entropy optimal policies.
Yuzheng HuZiwei JiMatus TelgarskyPublished in: ICLR (2022)
Keyphrases
- optimal policy
- average reward
- actor critic
- policy iteration
- reinforcement learning
- markov decision processes
- decision problems
- long run
- state space
- finite horizon
- partially observable markov decision processes
- infinite horizon
- policy gradient
- multistage
- initial state
- dynamic programming
- finite state
- sufficient conditions
- approximate dynamic programming
- optimal control
- average cost
- markov decision process
- model free
- temporal difference
- function approximation
- reinforcement learning algorithms
- fixed point
- search space