Data-driven traffic signal optimisation with random forest and ESO-based model-free adaptive iterative learning control

Urban traffic signal control plays a crucial role in optimising traffic flow, but traditional methods often struggle with large initial errors and limited adaptability to dynamic traffic conditions. This study presents a data-driven approach for traffic signal optimisation by integrating Random Forest (RF) with ESO-based Model-Free Adaptive Iterative Learning Control (MFAILC). Historical traffic data are utilised to train an RF model that generates optimised initial control signals, reducing initial iteration errors and accelerating convergence, while the MFAILC framework dynamically adapts control inputs using iterative learning and real-time traffic data. An Extended State Observer (ESO) is incorporated to estimate external disturbances and traffic uncertainties in the iteration domain, improving overall system performance. Simulation results in VISSIM demonstrate significant advantages on a 28-intersection network: under medium disturbance, ESO-RFMFAILC achieves an error standard deviation of 21.930 versus 24.580 for baseline MFAILC (10.4% improvement), with 20–25 iterations for convergence compared to 30–35 for baseline methods. The effectiveness is further examined across different network scales, showing consistent error reductions on small-scale networks (4 intersections) despite limited training data, and under varying disturbance conditions. We also investigate non-repetitive traffic conditions, where sudden pattern shifts induce noticeable error increases across all methods; nevertheless, ESO-RFMFAILC exhibits superior recovery behaviour compared with baselines. Overall, these findings suggest the proposed method is a promising direction for signal control in medium-to-large-scale networks with repeatable weekday traffic patterns under simulation settings, while practical deployability in highly dynamic, non-repetitive environments requires further field-oriented validation.