Develop continuous adaptation of the GPR coefficients

Develop continuous, gradient-based adaptation of the live-reward and pre-collected-reward coefficients α(t) and β(t) in the General Policy Reward objective, instead of the piecewise-constant batch-composition schedule used in the Multi-Agent Guided Reinforcement Learning protocol.

Background

The General Policy Reward objective combines live and pre-collected reward terms using coefficients α(t) and β(t). In the presented implementation, these coefficients are not learned continuously; they are induced by the composition of offline and online mini-batches across two training stages.

The authors identify continuous optimization of these coefficients as an unresolved extension. Such adaptation could allow the training objective to balance live and pre-collected experience autonomously throughout training, rather than changing only through a prescribed stage schedule.

References

This choice requires no additional tuning beyond the batch composition itself, at the cost of foregoing continuous, gradient-based adaptation of α,β, which we leave 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 5.3, subsection “MGARL: Design Rationale and Training Protocol,” paragraph “Piecewise-constant schedule for α,β”