Projected Off-Policy Q-Learning (POP-QL) for Stabilizing Offline Reinforcement Learning.
Melrose RoderickGaurav ManekFelix BerkenkampJ. Zico KolterPublished in: CoRR (2023)
Keyphrases
- reinforcement learning
- function approximation
- reinforcement learning algorithms
- state space
- optimal policy
- temporal difference
- action selection
- markov decision processes
- temporal difference learning
- continuous state and action spaces
- stochastic approximation
- model free
- multi agent
- learning algorithm
- query language
- reinforcement learning methods
- state action space
- multi agent reinforcement learning
- control problems
- real time
- learning agent
- learning process
- relational reinforcement learning
- eligibility traces
- partially observable
- state action
- transfer learning
- dynamic programming
- continuous state
- nonlinear systems
- multiple agents
- hierarchical reinforcement learning
- td learning
- database