Developing a learning hybrid algorithm for vehicle routing problem

In this thesis, an AI-supported, learnable, and hybrid optimization algorithm is developed to address vehicle routing problems commonly encountered in logistics and distribution operations. The proposed model offers an adaptive decision support system by integrating RL, neural network-assisted GA, and a NoSQL-based data processing infrastructure. Unlike traditional static algorithms, the developed model adapts to new scenarios by learning from past route outcomes. The model was tested on four real-world delivery scenarios conducted on both sides of Istanbul. Test results indicate improvements ranging from 18% to 37% in key operational metrics such as the number of vehicles, total distance, route duration, and shift time. The neural network-supported initial population helps mitigate the early convergence risk of the genetic algorithm, leading to better-quality solutions. The NoSQL integration enables fast and multi-criteria queries on large datasets, while the RL module allows continuous policy updates based on operational feedback. As a result, this study presents not only an alternative to traditional routing approaches but also a flexible, real-time, and field-learnable decision support solution, contributing significantly to the logistics sector.