Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Deep learning-guided Monte Carlo tree search for lane-free autonomous driving

Peridis Ioannis

Full record


URI: http://purl.tuc.gr/dl/dias/16741399-4329-4E65-BD4D-3D308CCB00B1
Year 2024
Type of Item Diploma Work
License
Details
Bibliographic Citation Ioannis Peridis, "Deep learning-guided Monte Carlo tree search for lane-free autonomous driving", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024 https://doi.org/10.26233/heallink.tuc.99857
Appears in Collections

Summary

Vehicular traffic management has become increasingly complex due to the rise in autonomous driving technologies. Traditional lane-based traffic systems, while structured and familiar, often struggle with optimization and dynamic flow control, leading to congestion and inefficiencies. By contrast, lane-free traffic environments offer a promising alternative by allowing vehicles to maneuver laterally across the entire roadway without the constraints of lanes, which could significantly enhance road capacity and traffic fluidity.This thesis explores the use of Monte Carlo Tree Search (MCTS) and supervised deep learning techniques to advance autonomous driving technologies. MCTS is well-suited for dynamic and unpredictable environments, such as autonomous driving, due to its robust decision-making capabilities in complex scenarios. Combined with Deep Neural Networks (DNNs), which excel in pattern recognition and predictive modeling from large datasets, these technologies could potentially revolutionize traffic management systems. A representation of MCTS in a Markov Decision Process (MDP) framework specifically tailored for lane-free traffic scenarios is developed, enhancing traditional MCTS approaches to better handle the demands of this environment. This thesis builds upon the existing MCTS model developed in a 2023 diploma thesis by Pantelis Giankoulidis,focusing initially on refining how the algorithm processes state information. Our enhancements significantly improved the operational efficiency and effectiveness of the MCTS framework, enabling it to handle high-density traffic situations more adeptly.Further advancements were achieved by integrating a neural network into the MCTS framework to guide the selection phase. This integration utilized the predictive capabilities of DNNs, allowing for more informed decision-making and faster exploration during the tree search process. Additionally, we investigated a standalone neural network approach, designed to function without the exploratory benefits of MCTS, to evaluate its comparative effectiveness in decision-making.Our thorough experimental evaluation demonstrates that the enhanced MCTS framework, supported by neural network guidance, markedly improves upon the lane-free vehicles’ behaviour. Key metrics such as safety, through reduced collision rates, and efficiency, by optimizing speed, highlighted the substantial benefits of this integrated approach. These results underscore the potential of combining MCTS with neural network technologies to aid the decision-making in autonomous vehicular driving environments.

Available Files

Services

Statistics