Artificial Intelligence and Graph Theory Based Traffic Signal Control Systems for Road Intersections

Abstract

Artificial Intelligence and Graph Theory Based Traffic Signal Control Systems for Road Intersections Traffic congestion is a critical and economically damaging problem in dense urban areas, driven by rapid urbanization, rising vehicle density, and — most critically — fixed-time signal strategies that fail to adapt to real-time demand. This work proposes a hybrid traffic signal control system that integrates fuzzy graph coloring (FGC) with deep reinforcement learning (DRL) for an isolated four-way road intersection. Fuzzy graph coloring is used to model uncertainty in traffic-flow data through membership functions, deriving adaptive, semantically structured traffic-phase groupings rather than the fixed, human-designed phases used in conventional DRL-based controllers. These fuzzy phase clusters — generated at three independence thresholds (α = 1.0, 0.92, 0.5) — serve as the action space for a Deep Q-Network (DQN) agent trained with prioritized experience replay, which learns to select and time signal phases to minimize cumulative vehicle delay. All experiments were conducted in the SUMO (Simulation of Urban Mobility) traffic simulator, with fuzzy computations performed in MATLAB and the DRL agent implemented in Python via the TraCI interface. Three DRL models (Model 1, Model 2, Model 3, corresponding to the three thresholds) were evaluated against a traditional fixed-time controller across six traffic-intensity levels (×0.25 to ×1.50 of the training baseline). Results show that the proposed models reduce waiting time by 6.27% to 59.09% relative to the traditional controller in lower-medium to higher-medium traffic intensities, with Model 2 (α = 0.92) achieving the best mean improvement (40.29%) and Model 1 (fully disconnected phases, α = 1.0) showing the most stable, scalable performance at high and extreme intensities. Effect-size analysis (Cohen's d > 2.5 for all three models) supports the practical significance of these improvements, while the accompanying statistical discussion transparently addresses the limitations of a small, non-replicated sample of intensity conditions. This repository contains the manuscript source and associated materials for the study. The key novelty lies in deriving the DRL action space itself from traffic-flow theory via fuzzy graph coloring, rather than using arbitrarily fixed phase groupings — a generalizable hybrid paradigm with potential applications beyond traffic management. Authors: M. Z. I. Bangalee, M. D. Ahmed, S. M. Tanim (Department of Applied Mathematics, University of Dhaka, Bangladesh) Keywords: Traffic Signal Control; Deep Reinforcement Learning; Fuzzy Graph Coloring; SUMO; Intelligent Transportation Systems