Data-Driven Algorithms For Resource Allocation Optimization

Abstract

In this thesis, we present a comprehensive solution for optimizing resource alloca- tion in an insurance company by integrating advanced machine learning techniques with integer programming. We start with an in-depth exploration of various machine learning models to predict key performance indicators (KPIs), focusing on both linear and non-linear regression models. Among these, decision trees and random forest models demonstrate the highest effectiveness due to their capability to manage com- plex, non-linear data relationships. These models provide accurate predictions for the number of incidents that each combination of motorcycles and platforms can handle, facilitating optimized deployment decisions. To enhance our solution, we incorporate a detailed analysis of edge cases and intro- duce additional parameters such as time of day and a custom "heaviness" metric. This ensures our models can adapt to diverse operational scenarios, providing re- liable predictions under varying conditions. Embedding these machine learning predictions into an integer programming framework allows us to optimize driver shift schedules, taking into account business constraints such as shift lengths and holidays. This approach minimizes costs while maximizing service efficiency, signif- icantly improving operational performance. Furthermore, we illustrate the advantages of using Optimization Programming Lan- guage (OPL) over the commonly used OR-Tools framework for solving complex ve- hicle routing and capacity management problems. Through a detailed case study, we demonstrate that OPL enhances model clarity and maintainability while opti- mizing performance in solving large-scale optimization problems. This comparison underscores the effectiveness of OPL in providing high-quality solutions for complex scenarios. Future work involves refining these machine learning models by exploring more advanced techniques like deep learning and reinforcement learning for greater pre- dictive accuracy and adaptability. Continued collaboration with industry partners is essential to validate and refine the models in various operational environments, ensuring the efficacy and applicability of our solution.