HIPODE: Enhancing Offline Reinforcement Learning with High-Quality Synthetic Data from a Policy-Decoupled Approach.
Shixi LianYi MaJinyi LiuYan ZhengZhaopeng MengPublished in: CoRR (2023)
Keyphrases
- synthetic data
- reinforcement learning
- high quality
- optimal policy
- policy search
- action selection
- markov decision process
- partially observable environments
- state and action spaces
- reinforcement learning problems
- real image data
- policy gradient
- markov decision processes
- control policies
- state space
- function approximators
- partially observable
- reward function
- control policy
- reinforcement learning algorithms
- approximate dynamic programming
- policy iteration
- action space
- model free
- data sets
- real world
- dynamic programming
- actor critic
- state action
- function approximation
- low quality
- model free reinforcement learning
- mri data
- machine learning
- decision problems
- control problems
- long run
- policy evaluation
- average reward
- learning problems
- markov decision problems
- learning process
- transition model
- high resolution
- image quality
- partially observable domains
- state dependent
- infinite horizon
- temporal difference
- real time
- neural network
- agent receives
- multi agent
- database
- inverse reinforcement learning
- continuous state spaces
- ground truth
- continuous state
- rl algorithms
- finite state
- partially observable markov decision processes
- agent learns
- reinforcement learning methods