This report summarizes and discusses the development, testing, and verification of a safety performance function and multi-objective optimization methodology using four principle algorithms that comprise the proposed adaptive system, Kadence, for tuning the cycle length, splits, offsets, and left-turn phase sequence and signalized intersections. The safety performance function was developed by training a cascade feed forward neural network to learn the relationship between the signal timing settings, efficiency measures, and the resulting average traffic conflict rates. The average traffic conflict rates were post-processed using SSAM from trajectory data from the VISSIM traffic simulation model. From training the network with approximately 150 scenarios, the average error on the prediction of traffic conflicts on cases that were not used for training was 17%. This average error represented a reasonable performance to use as the safety performance prediction function in the adaptive control system. The Virtual D4 controlled was used to implement the signal operation in VISSIM. MATLAB was used to implement the neural network and export the network calculations as a C++ DLL that was integrated with the rest of the adaptive system code.
Kadence performs best when all four algorithms (cycle length, splits, offsets, and left-turn phase) are enabled in the system parameters set-up. Enabling one algorithm alone, while disabling others, was found to be ineffective. Similarly, enabling one or multiple algorithms at only one intersection in a system was found to be in effective.
Kadence predicts the total number of conflicts associated with optimized signal timing settings, however more efficient timing settings did not always result in a lesser number of conflicts than pre-optimized signal timing settings.