Algorithmic analysis demonstrates sublinear regret in lost-sales inventory systems under unknown supply uncertainty, indicating robust optimization without prior distribution data.
Key Points
To develop an effective online learning algorithm for lost-sales inventory systems operating under stochastic lead times, random supplies, and unknown demand and supply distributions.
Formulated an inventory management model incorporating general supply uncertainty and censored feedback without relying on prior distribution knowledge.
Constructed an analytical framework based on transformed convexity combined with coupling and concentration inequalities to bound long-run costs.
Evaluated performance analytically via cumulative regret bounds against the best constant-order policy and verified performance using numerical simulations.
Derived the first provably effective learning algorithm for lost-sales inventory systems facing both stochastic lead times and random supply yields.
Proved a sublinear cumulative regret bound parameterized by the deterministic lead time component and the upper bound of the stochastic lead time variation.
Showed via numerical experiments that the learning algorithm consistently outperforms standard baselines under severe supply disruptions.