Expert Competitive Traffic Light Optimization with Evolutionary Algorithms

Abstract

We present a complete system to optimize traffic lights green phases and temporal offsets based on a combination of microscopic simulation and black box, evolutionary algorithms. We also report the outcome of an AI versus experts comparison workshop conducted with our algorithm and seasoned experts from a specialized traffic engineering office. Experimental results indicate that the proposed algorithmic scheme significantly outperforms expert efforts. Our system entails a memetic (genetic+gradient) calibration module to adapt the Origin/Destination (O/D) matrix to current traffic conditions, an inoculation procedure to incorporate existing traffic light programs, genetic multi-objective optimization capabilities and sound metrics. Experiments are conducted over several real world datasets of operational sizes from the Paris outskirts and various other French urban areas. Our experimental outcome is threefold. First, we report the success of the memetic calibration module in adjusting the simulator’s O/D matrix to a point with variation levels corresponding to recorded sensor data. Second, we confirm the ability of the system to obtain significant gains on that sound basis: gains ranging from 15% to 35% are consistently reached on both traffic jams reduction and pollutant emissions. Most importantly, we report the outcome of the comparison workshop: a formalized methodology followed by experts to manually optimize traffic lights, iterative experimental logs tracing the application of that methodology to two real world cases and comparable results obtained by the algorithm on the same cases. Results indicate that the AI module performs significantly better than experts in both speed and final solution quality.