Mémoires de Master

Permanent URI for this collectionhttps://dspace.univ-ghardaia.edu.dz.dz/handle/123456789/63

Browse

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Adaptive Traffic Signal Control Using Reinforcement Learning and Simulation
    (university of ghardaia, 2026) MEHAYA Fatima; ABDELHADI Imane
    Urban traffic congestion remains a critical challenge for modern cities, exposing the fundamental limitations of conventional fixed-time signal controllers. This thesis proposes and evaluates a fairness-aware reinforcement learning (RL)framework for adaptive traffic signal control, built around the Proximal Policy Optimisation (PPO) algorithm and implemented within the SUMO microscopic traffic simulator. A multi-objective reward function is designed to jointly optimise traffic efficiency and directional fairness, combining delay reduction, queue minimisation, throughput maximisation, and Jain’s Fairness Index. Three distinct weight configurations are evaluated: efficiencyoriented, fairness-oriented, and balanced. Five controllers are systematically compared FixedTime, Vanilla PPO, PPO Efficiency, PPO Fairness, and PPO Balanced across two traffic scenarios: a familiar Training Route (with static imbalanced demand) and an Unseen Route (with temporally-varying symmetric demand) for zero-shot generalisation assessment. Experimental results on the Training Route demonstrate that all PPO-based controllers reduce average waiting time by 40–47% relative to the Fixed-Time baseline. PPO Efficiency achieves the highest training throughput among RL controllers (655 vehicles) and the lowest waiting time (15.63 s). PPO Balanced provides the most consistent multi-objective performance, achieving the lowest average queue length (168.2 vehicles). On the Unseen Route, PPO Fairness demonstrates superior generalisation, retaining 130.7% of its training throughput (459 → 600 vehicles) while maintaining a Jain’s Index of 0.8314, well above the 0.8 acceptability threshold. Notably, all fairness-weighted controllers maintain acceptable directional equity under distribution shift. A key finding is that fairness-oriented reward shaping implicitly regularises the learned policy, improving robustness under distribution shift. These results offer practical deployment guidelines for smart city traffic management systems.