A neuro-dynamic programming approach to retailer inventory management

We discuss an application of neuro-dynamic programming techniques to the optimization of retailer inventory systems. We describe a specific case study involving a model with thirty-three state variables. The enormity of this state space renders classical algorithms of dynamic programming inapplicable. We compare the performance of solutions generated by neuro-dynamic programming algorithms to that delivered by optimized s-type ("order-up-to") policies. We are able to generate control strategies substantially superior, reducing inventory costs by approximately ten percent.