State Advantage Weighting for Offline RL
This work addresses offline RL challenges for AI researchers, offering a novel approach that improves performance and generalization, though it appears incremental as it builds on existing advantage concepts.
The paper tackles the problem of offline reinforcement learning by introducing state advantage weighting, which decouples actions from values using state advantage and QSS learning instead of action advantage and QSA learning. The result is remarkable performance on D4RL datasets and good generalization when transferring from offline to online settings.
We present state advantage weighting for offline reinforcement learning (RL). In contrast to action advantage $A(s,a)$ that we commonly adopt in QSA learning, we leverage state advantage $A(s,s^\prime)$ and QSS learning for offline RL, hence decoupling the action from values. We expect the agent can get to the high-reward state and the action is determined by how the agent can get to that corresponding state. Experiments on D4RL datasets show that our proposed method can achieve remarkable performance against the common baselines. Furthermore, our method shows good generalization capability when transferring from offline to online.