Dynamic Optimization Strategies for Multi-Modal Transit Systems

Abstract

The complexity of urban transportation networks in cities presents significant challenges for efficient management and optimization. These challenges are compounded by the diverse needs of different passenger groups, including often-neglected populations such as older peo ple. Furthermore, the intricate interactions between various modes of transportation, such as buses and subways, create a highly dynamic and interconnected system that is difficult to optimize holistically. Balancing the needs of all passengers while simultaneously addressing congestion and minimizing environmental impact adds another layer of complexity to this already multifaceted problem.

Current approaches to transit optimization often fall short in addressing the full scope of this challenge. Traditional methods in transit management typically use linear programming or genetic algorithms for bus route optimization and subway scheduling. However, these conventional approaches often rely on static, predetermined parameters and struggle to adapt to dynamic changes in demand or unexpected disruptions. They also tend to oversimplify complex urban dynamics, leading to suboptimal solutions in practice. In recent years, machine learning and deep learning techniques have been applied to improve transit optimization. For instance, some researchers have used neural networks for passenger flow prediction, while others have employed reinforcement learning for adaptive traffic signal control. However, many of these efforts still maintain a narrow focus on individual aspects of the system, such as route planning or demand prediction, and are often limited to single-modal transportation systems. Both traditional and ML-based approaches tend to prioritize overall system efficiency or congestion reduction, often overlooking the trade-offs between congestion management and environmental impact, as well as the specific needs of diverse passenger groups. There remain open research questions relating to the simultaneous optimization of multi-modal transit systems to balance congestion reduction with environmental concerns, while improving service for all passenger groups in the dynamic context of urban transportation networks.

This thesis addresses these limitations in transport optimization leveraging advanced AI tech niques. There are a number of challenges to be addressed, in particular: data imbalance in passenger classification, congestion prediction for proactive optimization, adaptation for pre diction models, and multi-objective optimization in multi-modal systems. To tackle the issue of data imbalance in passenger classification, generative AI models were employed. Gener ative Adversarial Networks (GANs) was explored to enhance the dataset and solve the data imbalance in passenger classification. For congestion prediction enabling proactive optimiza tion, a hierarchical algorithm was implemented, leveraging Graph Neural Networks (GNNs) and Transformer-based sequence models. This approach was further refined to a lighter pre dictive algorithm using Selective State Spaces, addressing computational complexity concerns while maintaining predictive power. To achieve efficient adaptation for prediction models, a ii dynamic ensemble framework for spatial-temporal forecasting was developed. This framework dynamically updates predictive models by periodically integrating new data, maintaining rele vance and precision as real-world conditions evolve. Finally, for multi-objective optimization in multi-modal systems, a multi-agent reinforcement learning model incorporating Kolmogorov Arnold Networks (KANs) was designed. This model integrates insights from the passenger classification and congestion prediction models to optimize transit schedules across various modes of transportation, balancing multiple objectives including operational efficiency and environmental considerations.

Simulation experiments were conducted using data sets from the multi-modal urban transit systems of Suzhou and Beijing. The models were evaluated using multiple metrics including standard metrics such as F-scores, Mean Squared Error (MSE), and Root Mean Squared Error (RMSE), as well as custom measures including passenger-type-based station congestion degree and daily CO2 emissions. The results demonstrated improvements for both regular commuters and typically neglected groups such as older people, with a reduction in overall daily congestion frequency and reduced pollution levels across the systems. The models effectively address multiple optimization objectives, enhancing both efficiency and environmental sustainability in integrated urban transit systems. However, it‘s important to note the limitations of this study. The passenger data was derived solely from AFC (Automated Fare Collection) systems, which may not fully capture all aspects of user behavior. Additionally, while the study covers three major Chinese cities, the model`s effectiveness in cities with significantly different urban layouts or transit systems may vary. Future work could focus on incorporating more comprehensive data sources for improved passenger behavior modeling and testing the model across a wider range of urban environments to enhance its generalizability. These steps would further validate and refine the model`s effectiveness in diverse urban contexts.