Optimal Convergence Rate for Exact Policy Mirror Descent in Discounted Markov Decision Processes.
Emmeran JohnsonCiara Pike-BurkePatrick RebeschiniPublished in: NeurIPS (2023)
Keyphrases
- markov decision processes
- discount factor
- convergence rate
- optimal policy
- policy iteration
- average reward
- learning rate
- average cost
- dynamic programming
- finite horizon
- discounted reward
- stationary policies
- infinite horizon
- total reward
- finite state
- partially observable
- state space
- markov decision problems
- markov decision process
- convergence speed
- policy iteration algorithm
- step size
- decision processes
- reinforcement learning
- transition matrices
- long run
- decision problems
- action space
- action sets
- state and action spaces
- state dependent
- decision theoretic planning
- reinforcement learning algorithms
- optimality criterion
- expected reward
- optimal control
- initial state
- continuous state spaces
- policy evaluation
- control policies
- reward function
- expected cost
- multistage