ConceptLatent Space Divergence
Latent Space Divergence
A measure of the distance between two distributions in a low-dimensional embedding space. In offline reinforcement learning it can serve as a theoretically bounded surrogate for the divergence of state-consequence distributions without requiring access to the true dynamics model.
Timeline (last 90 days)
Oct 30
The researchers propose Structural Behavior Regularization (StBR), which moves the constraint from action space to state-consequence space, uses latent-space divergence as a theoretically bounded surrogate objective, and adaptively tightens the behavior constraint in dynamically sensitive regions.
Unverified50%