A maximum-entropy approach to off-policy evaluation in average-reward MDPs.
Nevena LazicDong YinMehrdad FarajtabarNir LevineDilan GörürChris HarrisDale SchuurmansPublished in: CoRR (2020)
Keyphrases
- maximum entropy
- average reward
- policy evaluation
- markov decision processes
- policy iteration
- model free
- optimal policy
- reinforcement learning
- partially observable markov decision processes
- finite state
- markov models
- temporal difference
- state action
- markov decision process
- least squares
- fixed point
- state space
- reinforcement learning algorithms
- dynamic programming
- average cost
- markov decision problems
- conditional random fields
- function approximation
- linear programming
- reward function
- long run
- action space
- monte carlo
- pairwise
- policy gradient
- optimal control
- partially observable
- decision problems
- supervised learning
- infinite horizon