Simulation times Vs. network size in a genetic algorithm based urban traffic optimization architecture

Abstract

Our group has developed and successfully tested a new urban traffic lights optimization architecture. This architecture includes a standard Genetic Algorithm, a Cellular Automata based traffic simulator within the evaluation function, and a Beowulf Cluster. Every hypothetical solution is marked according to the statistics obtained from a traffic micro-simulation. So far, the duration of each simulation has been chosen without any formal restriction. With this paper, we want to share the first results and conclusions of our study regarding the simulation times against the traffic network size comparison. To do so, we have run a wide set of experiments with different scaled networks. For this paper, we have presented the results obtained with a small "lab network" and a big network taken from the real world - "Las Ramblas" in Santa Cruz de Tenerife, Spain. We have found a strong linear relationship among simulation times and fitness. It could mean big computing time savings for our system. This is particularly interesting for a future adaptation of the model to a dynamic environment, with real time restrictions.