Learning to Undo: Transfer Reinforcement Learning under State Space Transformations
An “undo” map lets an RL agent reuse a learned policy with fewer samples than training from scratch, under the paper’s stated assumptions.
The method matches state-feature statistics from source and target environments to learn a mapping that recovers the source’s state space. It then uses that map to transfer the source policy. The sample-complexity guarantee applies when the map is linear and the source’s optimal action-value function is representable with linear features. The authors also report improved sample efficiency on continuous-control tasks and Atari games beyond that theoretical setting.
ArXiv · AI/CL/LG's note
The method matches state-feature statistics from source and target environments to learn a mapping that recovers the source’s state space. It then uses that map to transfer the source policy. The sample-complexity guarantee applies when the map is linear and the source’s optimal action-value function is representable with linear features. The authors also report improved sample efficiency on continuous-control tasks and Atari games beyond that theoretical setting.
ArXiv · AI/CL/LG's note
score 4