Metaheuristic Optimization for Enhancing Performance of Predictive Vehicle Control Delay Estimation Model

Efficient traffic signal management relies on precisely anticipating vehicle control delays, critical for improving traffic flow. Techniques for predicting vehicle control delay have improved, especially since artificial intelligence (AI) and artificial neural networks (ANNs) were introduced. However, the inherent complexity of ANN and other gradient-based algorithms presents difficulties due to numerous local minima in each layer, preventing optimal solutions. Furthermore, the complex dynamics of mixed traffic require thorough parameter adjustment to improve responsiveness and precision. Motivated by these problems, this study incorporates nature-based techniques into artificial neural networks to enhance the accuracy of vehicle control delay estimate models. The hybrid system, genetic ant bee particle swarm grey wolf simulated annealing-artificial neural network (GAPWS-ANN), uses five nature-inspired optimization algorithms [genetic algorithm (GA), ant bee colony (ABC), particle swarm optimization (PSO), grey wolf optimization (GWO), and simulated annealing (SA)] to fine-tune the neural network’s hyperparameters successfully. When applied to predicting vehicle control delays in an isolated intersection under mixed traffic conditions, GAPWS-ANN outperformed the base ANN model through thorough comparison and mathematical validation. Extensive data from video and on-site surveys at the Shyamal crossroads contributed to a complete database, which included geometric parameters and vehicle-related information. The hybrid models outperformed the baseline ANN model, with GA-ANN and PSO-ANN achieving remarkable results in both the training and testing stages. The analysis showed significant improvements, with R2 values of 0.9949 and 0.9967 during training and 0.9915 and 0.9918 during testing, respectively. Furthermore, their RMSE values showed improved accuracy, measuring 2.94 and 2.89 during training and 3.6154 and 3.5740 during testing, respectively.