Reinforcement Learning with Logarithmic Regret and Policy Switches.
Grigoris VelegkasZhuoran YangAmin KarbasiPublished in: NeurIPS (2022)
Keyphrases
- reinforcement learning
- reward function
- total reward
- optimal policy
- policy search
- worst case
- markov decision processes
- markov decision process
- action selection
- reinforcement learning algorithms
- state space
- regret bounds
- action space
- partially observable
- partially observable environments
- policy iteration
- reinforcement learning problems
- average reward
- markov decision problems
- function approximation
- policy evaluation
- model free
- state and action spaces
- inverse reinforcement learning
- actor critic
- policy gradient
- state action
- control policies
- function approximators
- partially observable markov decision processes
- long run
- dynamic programming
- approximate dynamic programming
- lower bound
- reward signal
- control policy
- expected reward
- multi agent
- multi armed bandit problems
- partially observable domains
- finite horizon
- expert advice
- least squares
- regret minimization
- infinite horizon
- multiple agents
- minimax regret
- machine learning
- continuous state spaces
- continuous state
- agent learns
- finite state
- decision problems
- transition model
- reinforcement learning methods