Neuro-Fuzzy Green-Time Allocation for Oversaturated Four-Way Signalized Intersections

Abstract

Fixed-time traffic signals allocate a constant green duration regardless of how vehicle demand fluctuates across approaches. That rigidity becomes critical once an approach enters oversaturation, that is, once its degree of saturation x, the ratio of the arrival rate per cycle to the capacity the signal makes available per cycle, reaches or exceeds unity, so that a residual queue is carried over from one cycle to the next and successive residuals accumulate. This study designs and simulates an adaptive green-time allocation controller based on an Adaptive Neuro-Fuzzy Inference System (ANFIS) with two inputs, aggregate phase queue length (Q) and maximum approach waiting time (W), for a four-way intersection operated from undersaturated to oversaturated conditions. The training data are not arbitrary: they are generated from a queue-actuated green-time formula whose three terms, queue discharge time, startup lost time and an anti-starvation priority proportional to normalized waiting time, are individually grounded in signal-timing theory, and whose parameters are reported in full. The model, using three Gaussian membership functions per input, nine first-order Sugeno rules and hybrid learning, was trained and validated on an 80:20 split of 180 data pairs and reached a testing RMSE of 2.39 s and an MAE of 1.87 s, that is, 2.1% of the 115-second output range. Performance was then evaluated over a 100-cycle discrete-event simulation with Poisson arrivals against fixed-time control at a constant 30 s green, using an identical queue-update mechanism. Under the baseline plan the effective capacity is 7 vehicles per approach per cycle, and the three scenarios in which at least one approach exceeds it (x = 1.14) are exactly the scenarios in which the fixed-time queue grows linearly with a strictly positive drift and in which the proposed controller yields improvements of 68.6% to 87.1%. In the four scenarios in which no approach exceeds capacity the improvement is 0.8% to 10.0%. The benefit of neuro-fuzzy green-time allocation is therefore concentrated in, and structurally explained by, the oversaturated regime.