How AI is Redrawing City Maps?

Imagine being able to predict where people will go in a city without asking anyone. Thanks to data from smartphones and smart algorithms, researchers are developing models that simulate human movement with high precision. This study explores three advanced techniques — SPSA, PSO, and ADAM — to calibrate these models using 'opportunistic' data, such as that collected passively from apps or mobile devices. The goal? Make traffic and transport forecasts more accurate even with limited data. ADAM, known in AI, is applied here for the first time to this kind of problem. The results show it's possible to improve urban mobility simulations, making them more efficient and less dependent on costly surveys. A step forward for smarter, less congested cities.

This study focuses on the aggregate calibration of destination choice models using opportunistic data, an increasingly relevant source in human-centered contexts. Travel demand models are calibrated through simulation-based optimization, formulating the problem as the minimization of an objective function comparing the observed matrix (derived from external sources) with the simulated one. Three algorithms are compared: SPSA (Simultaneous Perturbation Stochastic Approximation), PSO (Particle Swarm Optimization), and ADAM (Adaptive Moment Estimation). SPSA is an iterative method that uses stochastic estimates to optimize complex functions with few parameters. PSO mimics the collective behavior of bird flocks, updating participant positions in the solution space. ADAM, on the other hand, combines advantages of adaptive gradient-based methods, automatically adjusting learning rates. The application of ADAM in this context represents a novelty compared to existing literature. The experimental environment uses an agent-based model with a nested structure, implemented on a small real-world network. The tested objective functions evaluate the goodness of fit between observed and simulated matrices, considering metrics such as mean squared error and Kullback-Leibler divergence. The results indicate that ADAM achieves competitive performance, especially in terms of convergence speed, though it shows some limitations in extreme scenarios. The study opens new avenues for the use of evolutionary algorithms in transport modeling, particularly in data-scarce environments.