Revisiting the application of simultaneous perturbation stochastic approximation towards signal timing optimization

Simultaneous Perturbation Stochastic Approximation (SPSA) has gained favor as an efficient optimization method for calibrating computationally intensive, “black box” traffic flow simulations. Few recent studies have investigated the efficiency of SPSA for traffic signal timing optimization. It is important for this to be investigated, because significant room for improvement exists in the area of signal optimization. Some signal timing methods and products perform optimization very quickly, but deliver mediocre solutions. Other methods and products deliver high-quality solutions, but at a very slow rate. When using commercialized desktop signal timing products, engineers are often forced to choose between speed and solution quality. Real-time adaptive control products, which must optimize timings within seconds on a cycle-by-cycle basis, have limited time to reach a high-quality solution. The existing literature indicates that SPSA provides the potential for upgrading both off-line and on-line solutions alike, by delivering high-quality solutions within seconds. This paper describes an extensive set of optimization tests involving SPSA and genetic algorithms (GA). The final results suggest that GA was slightly more efficient than SPSA. Moreover, the results suggest today’s signal timing solutions could be improved significantly by incorporating GA, SPSA, and ‘playbooks’ of pre-optimized starting points. However, it may take another 5-10 years before our computers become fast enough to simultaneously optimize coordination settings (i.e., cycle length, phasing sequence, and offsets) at numerous intersections, using the most powerful heuristic methods, at speeds that are compatible with real-time adaptive solutions