AACL learns a shared action representation via inverse dynamics so a policy can transfer across changing action spaces, and outperforms popular continual RL baselines on MiniGrid, Procgen Bigfish, and Atari Atlantis benchmarks.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Action-Adaptive Continual Learning: Enabling Policy Generalization under Dynamic Action Spaces
AACL learns a shared action representation via inverse dynamics so a policy can transfer across changing action spaces, and outperforms popular continual RL baselines on MiniGrid, Procgen Bigfish, and Atari Atlantis benchmarks.