Contextual Policy Transfer in Meta-Reinforcement Learning via Active Learning.
Jingchi JiangLian YanXuehui YuYi GuanPublished in: WISA (2022)
Keyphrases
- active learning
- transfer learning
- reinforcement learning
- optimal policy
- policy search
- markov decision process
- learning algorithm
- action selection
- policy iteration
- policy gradient
- knowledge transfer
- reinforcement learning problems
- contextual information
- markov decision processes
- partially observable environments
- supervised learning
- partially observable
- reinforcement learning algorithms
- active exploration
- markov decision problems
- function approximators
- state space
- learning process
- state and action spaces
- function approximation
- actor critic
- meta level
- policy evaluation
- labeled data
- reward function
- control policies
- action space
- previously learned
- control policy
- average reward
- state action
- model free
- approximate dynamic programming
- exploration exploitation
- state dependent
- selective sampling
- dynamic programming
- relevance feedback
- learning tasks
- optimal control
- temporal difference
- random sampling
- long run
- decision problems
- transferring knowledge
- cross domain
- continuous state
- active learning strategies
- partially observable domains
- machine learning
- semi supervised learning
- semi supervised
- partially observable markov decision processes
- decision trees
- multi agent
- transition model
- training set
- learning strategies
- agent learns
- context sensitive
- pool based active learning
- batch mode
- training examples
- finite state
- cost sensitive
- learning problems