Meta-Q-Learning
meta-reinforcement-learningq-learningoff-policycontinuous-control
Abstraction: Off-policy meta-RL via Q-learning with trajectory context variable
Key points:
- Meta-Q-Learning (MQL) is an off-policy meta-RL algorithm presented at ICLR 2020
- Q-learning with a context variable (representation of past trajectory) matches state-of-the-art meta-RL if given that context
- Multi-task objective maximizes average reward across training tasks as effective meta-training signal
- Propensity estimation applied to recycle past replay buffer data for off-policy adaptation on new tasks, amplifying available data
- Evaluated on standard continuous-control benchmarks; compares favorably to prior meta-RL methods
Connections: Meta Reinforcement Learning · Q Learning · Off Policy Learning