Answers are grounded in this paper and cite their sources.
Capacity-limited inventory networks must reconcile locally requested shipments with executable warehouse flows while controlling losses under supply and demand disruptions. We introduce a decentralized, parameter-shared policy trained through a differentiable, mass-balanced simulator. A nodewise command-to-realization residual, derived from discrete inventory conservation, penalizes infeasible requested flows. An online-adjusted risk weight couples mean episode cost with the upper-tail conditional value-at-risk (CVaR). Ten independent seeds compare seven policies across four synthetic disturbance regimes (400 held-out trajectories per method, seed and regime). Under combined stress, mean cost is 4906.8 versus 4979.7 for risk-neutral policy optimization, while CVaR90 is 8464.8 versus 8580.9. Against a fixed-risk comparator with a closely matched final risk weight, the adaptive formulation further reduces stress CVaR90 by 39.4 cost units. Requested-but-unshipped flow drops from 15.80 to 2.27 units/day. Realized transitions conserve inventory under every policy: the residual regulates planning consistency, not physical conservation itself. Results reflect a reproducible synthetic benchmark rather than field validation.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: