Variance-Aware Regret Bounds for Undiscounted Reinforcement Learning in MDPs.
Mohammad Sadegh TalebiOdalric-Ambrym MaillardPublished in: CoRR (2018)
Keyphrases
- markov decision processes
- reinforcement learning
- multi armed bandit
- regret bounds
- markov decision problems
- policy iteration
- average reward
- optimal policy
- reinforcement learning algorithms
- state space
- markov decision process
- finite state
- state and action spaces
- infinite horizon
- lower bound
- partially observable
- function approximation
- action space
- online learning
- stochastic games
- model free
- dynamic programming
- continuous state and action spaces
- learning algorithm
- reward function
- action selection
- covariance matrix
- upper bound
- average cost
- function approximators
- optimal control
- learning problems
- machine learning
- partially observable markov decision processes
- probability distribution