Contextual policy transfer in reinforcement learning domains via deep mixtures-of-experts.
Michael GimelfarbScott SannerChi-Guhn LeePublished in: UAI (2021)
Keyphrases
- reinforcement learning
- transfer learning
- optimal policy
- partially observable domains
- policy search
- cross domain
- transferring knowledge
- markov decision process
- control policy
- action selection
- partially observable
- state space
- function approximation
- contextual information
- state and action spaces
- markov decision processes
- markov decision problems
- control policies
- function approximators
- actor critic
- policy gradient
- reinforcement learning problems
- complex domains
- reinforcement learning algorithms
- action space
- policy evaluation
- application domains
- model free
- knowledge transfer
- policy gradient methods
- dynamic programming
- continuous state spaces
- context sensitive
- target domain
- real world
- state dependent
- policy iteration
- infinite horizon
- reinforcement learning methods
- partially observable environments
- mixture model
- approximate dynamic programming
- learning tasks
- transition model
- temporal difference
- active learning
- deterministic domains
- partially observable markov decision processes
- decision problems
- blind source separation
- context dependent
- continuous state
- reward function
- rl algorithms
- state action
- average reward
- finite state