Attention-based Partial Decoupling of Policy and Value for Generalization in Reinforcement Learning.
Nasik Muhammad NafiCreighton GlasscockWilliam H. HsuPublished in: ICMLA (2022)
Keyphrases
- reinforcement learning
- optimal policy
- markov decision process
- policy search
- action selection
- policy iteration
- markov decision processes
- partially observable environments
- state and action spaces
- state space
- input output
- policy gradient
- control policy
- action space
- control policies
- reinforcement learning algorithms
- reward function
- function approximation
- state action
- markov decision problems
- autonomous mental development
- learning algorithm
- control problems
- function approximators
- reinforcement learning problems
- partially observable
- actor critic
- approximate dynamic programming
- policy evaluation
- multi agent
- dynamic programming
- long run
- continuous state
- model free
- partially observable markov decision processes
- state dependent
- transition model
- finite state
- temporal difference learning
- average reward
- machine learning
- model free reinforcement learning
- exploration exploitation tradeoff