dc.description.abstract | This thesis tries to address the logistics strategy selection problem in distribution centers and transportation considering two main criteria pertaining to the decision: product characteristics and customer satisfaction. We tried to analyze the effect of product characteristics, lead time, and fill rate on the logistics strategy selection. The primary purpose of the thesis is to develop a tool with the capability of classifying products with different characteristics to their suitable logistics strategy.
Due to data limitations, we first try to utilize the distribution-free bootstrap method to augment the data information and develop products with new characteristics. This is done with the help of bootstrap estimation of different fractiles of each data feature. After the data augmentation, we use a more comprehensive method to generate data based on learning of the underlying distribution of data in feature space by employing the Wasserstein Generative Adversarial Network with gradient penalty (W-GAN). This method has emerged to tackle small sample size data for generative adversarial networks. The WGAN-GP used in the thesis is constructed from a three-layer generator and discriminator with dropout later and activation function of tan h (hyperbolic tangent).
Following that, we construct a multi-layer perceptron architecture network for the classification of the new products with different characteristics. The sigmoid activation function of the constructed neural network transforms the network's binary label to a continuous label of between zero and one, enabling us to employ the result as an input of a utility function. After training the neural network, different utility functions, considering lead time, fill rate, and product characteristics, are defined. Linear, quadratic, square root and logarithmic utility functions are described and compared.
In this thesis, we started with a sample data size of 20 and enriched the data by incorporating field knowledge such as extreme point analysis and including single feature impact. Following that, we applied the bootstrapping method to enlarge the sample size. Finally, a W-GAN network was used to decode the underlying feature distribution and enables us to generate more samples for training the classifier. We could obtain 82% accuracy on training data and 80% on the test data which is quite a significant result given the coupling between the features. | en |