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