An intelligent Geographice Information System for Vehicle Routing Toka Sabry Mostafa

Supervisor: Hoda Talaat

Thesis (M.A.)—Nile University, Egypt, 2010 .

"Includes bibliographical references"

Contents:
The Vehicle Routing Problem with Time Window (VRPTW)
2.2. Geographic Information Systems (GIS
2.3. Global Positioning Systems (GPS)
2.4. Reinforcement Learning (RL)
3. IGIS-CVR MODEL FRAMEWORK
3.1. Model description
3.2. System Workflow
3.3. Model Architecture
3.4. The GIS Model
3.4.1. System’s Functionalities
4. THE RIENFORCMENT LEARNING MODEL
4.1. RL Principles
4.2. RL Basic Elements mapped to IGIS-CVR
4.3. Solution Approach
4.4. Elements of Q-Learning
4.4.1. Exploration and Exploitation
4.4.2. Learning Rate
4.4.3. Discount Factor
4.5. RL Model Implementation
5. SIMULATION BASED EVALUATION MODEL
5.1. Simulation Environment
5.2. Case study and Prototype implementation
5.2.1. Network definition
5.2.2. Mapping case study to RL elements
5.3. RL Prototype Implementation
5.4. Preliminary Analysis
5.4.1. Learning rate (α)
5.4.2. Training life time estimation
5.5. Case Study Assessment
5.5.1. Case 1: Static Case Evaluation
5.5.2. Case 2: Dynamic Case Evaluation

Abstract:
In Egypt, freight movement relies heavily on road transport. Commercial vehicles (CVs) constitute a
major segment of the vehicle population that travels the country’s roads contributing to (and suffering
from) daily congestion. Enhancing CVs operations by minimizing their en-route travel times benefits both
traffic network users as well as CV's business owners. The absence of traffic data collection
infrastructure, in many developing countries, impedes the usage of readily available vehicle routing
systems. In this research, we introduce a full modeling framework of an Intelligent Geographical
Information System for Commercial Vehicle Routing (IGIS-CVR). IGIS-CVR integrates a Geographic
information system (GIS) and a Reinforcement learning (RL) system to address the Vehicle Routing
Problems with Time Windows (VRPTW). The developed model uses CVs as probe vehicles for on-themove
data collection. Collected data is manipulated through a self-adaptive learning environment to
capture traffic network dynamics.
The scope of this research is focused on developing and implementing a prototype for the
system’s learning model. The Q-learning concepts of the Temporal Difference (TD) solution approach of
RL are used in the model formulation. Different estimation procedures for the model main parameters are
explored. A simulation based evaluation model is introduced to test the performance and effectiveness of
the developed RL prototype in two cases; 1) a static case under typical traffic conditions, 2) a dynamic
case where changes in network conditions are imposed. The evaluation results of the prototype highlight
its potential in effectively enhancing CV routing operations. An average decrease in the percentage of late
arrivals by; 77.5% for the static case, and 48% for the dynamic case are achieved through the model
implementation.

Text in English, abstracts in English .