Learning from the aggregated optimum: Managing port wine inventory in the face of climate risks
研究将波特酒库存管理建模为马尔可夫决策过程，通过聚合酒龄类别并利用机器学习训练决策规则，在气候风险下优化采购、生产和发行决策，相比基准规则年利润提升21.4%。
Port wine stocks ameliorate during storage, facilitating product differentiation according to age. This induces a trade-off between immediate revenues and further maturation. Varying climate conditions in the limited supply region lead to stochastic purchase prices for wine grapes. Decision makers must integrate recurring purchasing, production, and issuance decisions. Because stocks from different age classes can be blended to create final products, the solution space increases exponentially in the number of age classes. We model the problem of managing port wine inventory as a Markov decision process, considering decay as an additional source of uncertainty. For small problems, we derive general management strategies from the long-run behavior of the optimal policy. Our solution approach for otherwise intractable large problems, therefore, first aggregates age classes to create a tractable problem representation. We then use machine learning to train tree-based decision rules that reproduce the optimal aggregated policy and the enclosed management strategies. The derived rules are scaled back to solve the original problem. Learning from the aggregated optimum outperforms benchmark rules by 21.4% in annual profits (while leaving a 2.8%-gap to an upper bound). For an industry case, we obtain a 17.4%-improvement over current practices. Our research provides distinct strategies for how producers can mitigate climate risks. The purchasing policy dynamically adapts to climate-dependent price fluctuations. Uncertainties are met with lower production of younger products, whereas strategic surpluses of older stocks ensure high production of older products. Moreover, a wide spread in the age classes used for blending reduces decay risk exposure. • Periodic Markov decision process formulation for port wine inventory management, integrating purchasing, production, and flexible issuance decisions. • Analysis of the impact of climate-dependent purchase price risks shows that producers may even benefit from increased volatility. • Large-scale original problems are solved by tree-based rules trained on the optimal policy of an aggregated problem. • Upper bound on average reward based on optimization of average system state provides reference for scaling tree-based rules back to original problems.