Hybrid Neural Network and Genetic Algorithm Approach for Adaptive Traffic Signal Timing Optimization

Abstract

Urban traffic congestion remains one of the most pressing challenges in modern cities, causing significant losses in commute time, fuel consumption, and air quality. Traditional fixed-cycle signal control systems fail to adapt to the dynamic, unpredictable nature of urban traffic, leading to increased intersection delays and network-wide inefficiencies. This paper proposes a two-stage hybrid soft computing framework that combines a Multilayer Perceptron (MLP) Neural Network for multi-variate traffic volume prediction with a PyGAD-based Genetic Algorithm (GA) for adaptive signal timing optimization. In the first stage, the MLP model is trained on a synthetic dataset comprising 8,640 instances across four intersection archetypes — commercial, residential, business, and industrial — using 18 diverse input features spanning temporal, environmental, infrastructure, and historical lag categories. The MLP achieves a test R² value of 0.887, confirming strong generalization with minimal overfitting. In the second stage, the GA employs a threegene chromosome encoding green times and yellow duration, optimized using a composite fitness function with penalty terms to enforce realistic cycle constraints. Three representative urban scenarios — morning peak at a commercial intersection under rain, evening peak at a business intersection under clear conditions, and a weekend afternoon residential scenario — are evaluated. The proposed system achieves a consistent 20–25% reduction in average vehicle waiting time compared to conventional fixed-time signal control, demonstrating practical potential for deployment in Intelligent Transportation Systems (ITS).