Applications of Machine Learning in Revenue Management and Routing

Abstract

In this thesis, I use machine learning techniques to solve issues in revenue management and public transportation design. The first two chapters relate to problems of revenue management and online learning. The problem of sequential learning and optimization of the demand function has been an important topic in revenue management. Finding the optimal policy faces numerical complexity and is prone to the curse of dimensionality. It is mostly solved using heuristics and restrictive assumptions. In the first chapter, I use a novel non-parametric approach to solving dynamic pricing and learning problems. I develop a flexible method to approximate the optimal policy using polynomial approximation, thus reducing complexity. I make use of the Bayesian framework to update the probability model and make advances in numerical methods to solve this problem. In the second chapter, I use a machine learning heuristic called Thompson sampling. I improve the performance of the heuristic over short horizons by enforcing the decreasing nature of the demand function in the sampling algorithm. Using a stylized proof, I demonstrate the performance gains associated with this method and show the merits of ordered sampling with Thompson Sampling over short horizons. The last chapter makes use of a machine learning approach called data envelopment analysis (DEA), which I use in designing new public transportation routes in rural regions. I develop algorithms and heuristics to balance cost and equity under multiple objectives. The solution to this project was implemented in the city of Quinte West, Ontario.