Ablate the Anchored Transfer Stabilization components

Systematically ablate the residual behavior-cloning floor λ_res, the SymLog reward transformation, and the frozen Z-score normalization to isolate their individual contributions to the stability and performance of the Anchored Transfer Stabilization Protocol.

Background

The Anchored Transfer Stabilization Protocol combines three mechanisms intended to stabilize the offline-to-online transition: a residual behavior-cloning regularization floor, a Symmetric Logarithmic transformation of reward targets, and frozen input-normalization statistics obtained from the offline dataset.

The paper reports the combined framework but does not separately quantify the effect of each component. An ablation study is explicitly deferred, leaving unresolved which mechanisms are responsible for particular gains in stability, sample efficiency, or final reliability.

References

A systematic ablation isolating the individual contributions of the Anchored Transfer components---the residual BC floor $\lambda_{\mathrm{res}$, the SymLog reward transform, and the frozen Z-score normalization---is left to future work.

— From Prior-Guided Heuristics to Deployable Agents: Accelerating Demonstration-Driven Reinforcement Learning for Deadline-Constrained Network Control  (2609.03590 - Vitale et al., 3 Sep 2026) in Section Conclusions and Future Directions

A further open question concerns the sensitivity of \ac{mgarl} to the Stage-1/Stage-2 transition point $T_1$: a premature transition may leave the Actor and Critic insufficiently converged on $\mathcal{D}{\mathrm{off}$, exacerbating the Actor-Critic Misalignment discussed in Section~\ref{sec:mgarl_design_rationale}, whereas an excessively long pre-training phase may over-anchor the Actor to $\mu{\mathrm{MB}$, compounding the restrictive-anchor effect empirically observed under the Abilene $L=5$ configuration (Section~\ref{sec:results_rl_based_approaches}). A systematic study of $T_1$---jointly with the residual floor $\lambda{\mathrm{res}$ and the Early Stopping criterion that determines it in practice---is left to future work.

— From Prior-Guided Heuristics to Deployable Agents: Accelerating Demonstration-Driven Reinforcement Learning for Deadline-Constrained Network Control  (2609.03590 - Vitale et al., 3 Sep 2026) in Section Conclusions and Future Directions