A simple hidden-target-and-project strategy is provably optimal for inventory optimization with memory constraints, and viewing inventory as a one-dimensional queue dramatically simplifies the theoretical analysis.
This paper solves online inventory optimization—a practical problem where past inventory decisions constrain future actions—by maintaining a hidden target and projecting it onto feasible inventory levels. The method achieves optimal regret bounds on general convex capacity constraints, improving prior results and introducing a novel 'norm alignment' principle that simplifies the analysis.