Deriving interpretable decision rules from optimal purchasing and blending policies in port wine inventory management

Tactical port wine management is characterized by two decision periods per year. Purchasing decisions after the grape harvest and sales decisions after having reassessed the quality of stocks in January both have to factor in the current state of the multi-age inventory system. Flexibility in sales is provided through the possibility of blending stocks from different age classes to fulfill the demand for a certain target age. Further, decision-makers deal with stochastic harvest-dependent purchase prices as well as the risk of quality decay of stocks. We propose an infinite-horizon periodic Markov decision process model to integrate these uncertainties and solve it via value iteration for a small case study. Our findings indicate that widespread age patterns are favored for blending younger wine products which facilitates keeping strategic surpluses of older stocks. In order to derive interpretable decision rules for large-scale problems, we employ decision tree learning. When comparing this approach to naïve decision rules, initial simulation results show a significant increase in average profits.     «

Tactical port wine management is characterized by two decision periods per year. Purchasing decisions after the grape harvest and sales decisions after having reassessed the quality of stocks in January both have to factor in the current state of the multi-age inventory system. Flexibility in sales is provided through the possibility of blending stocks from different age classes to fulfill the demand for a certain target age. Further, decision-makers deal with stochastic harvest-dependent purcha...     »