SPAR: Support-Preserving Action Rectification

A new framework called Support-Preserving Action Rectification (SPAR) has been proposed to address challenges in offline policy improvement, particularly the conflict between maximizing value and fitting data distribution. SPAR utilizes a local residual rectification approach anchored to a frozen behavior cloning policy, enabling fine-grained fitting and local policy improvement in the residual space.

This development is significant as it introduces a mechanism, Latent Self-Imitation, which aims to resolve fitting-improvement gradient conflicts, potentially enhancing the effectiveness of AI models in decision-making tasks.