Deep Learning Model for Approximation of Inventory Costs under Stochastic Demand

Inventory control is one of the key aspects of organisational management for most businesses in today’s technologically advanced world. Finding the solution to optimise inventory control for different periods has been widely studied, and different statistical models have been constructed to tackle relevant problems. Yet, building up the most efficient model has been a challenge for the researchers as certain factors such as the usage of resources and the time of generating the optimal policy limit the models to a high extent. This paper suggests an approximation method to solve a single-item, single-location inventory problem under demand uncertainty for four periods by approximating inventory costs, incorporating a dynamic programming technique. The non-stationary uncertain demand was modelled using Poisson distribution, and six different demand patterns were incorporated. A stochastic Dynamic Programming model was used to generate the data for the modelling. The first step included the optimisation of the cost function of the dynamic programming to reduce the computational time. Subsequently, 5,431 instances of data were generated and fed into the approximation model. A Deep Learning algorithm - Recurrent Neural Networks – was built up for forecasting inventory costs. The configuration space of the hyperparameters was formulated, and the model was fine-tuned using the Random Search optimisation method. All the coding and modelling were carried out using Python programming language in Microsoft Visual Studio Code software. The Deep Learning model was proposed with a 12.98% MAPE score. Further suggestions and recommendations were provided to perfect the model for future use.

To request an access, email ezeynalli@hotmail.com.