Here are some key findings of this study:
A surrogate-assisted evolutionary algorithm can be combined with a local search strategy to enhance its anytime behaviour in traffic signal optimiza- tion problems.
Approximation models can be utilized to construct surrogate models which can be used to estimate fitness values of candidate solutions in traffic signal optimization using multi- objective evolutionary algorithms. Solutions which are evaluated by a traffic simulator during the optimization process are archived in a database to construct and update the surrogate model. This surrogate is able to learn the relationship between the inputs which are phase durations and the outputs that are traffic parameters needed such as traffic flow and delay. In each generation of the optimization process, new traffic simulator-based solutions are added into the database in order to update the surrogate model in order to improve the approximation accuracy in the oncoming iterations.
A local search can be utilized inside every generation of a surrogate-assisted evolutionary algorithm to improve anytime behaviour of the optimization algorithm in traffic light signal control systems. Some candidate solutions will be estimated by the surrogate model while the fitness value of the other solutions is evaluated by the traffic simulator.
A management model of the surrogate needs to be introduced to determine which will be used to estimate the fitness value of a candidate solution, the surrogate model or the traffic simulator. By using the surrogate model in assessing the goodness of solutions, the number of traffic simulation-based evaluations in each generation of the optimization process is reduced. As a result, using the same number of traffic simulation-based evalu- ations, the surrogate-assisted optimization algorithm has a larger number of generations than the non-surrogate algorithm. Therefore, anytime behaviour of the surrogate-based optimization algorithm will be enhanced. The local search is utilized in every itera- tion of the optimization process to accelerate the convergence rate. As a result, this combination improves anytime behaviour of the evolutionary algorithm in traffic signal optimization problems.
In traffic signal optimization problems, a local search method can be in- tegrated inside the iteration process of evolutionary algorithms to improve anytime behaviour.
This study proposed a local search method to improve anytime behaviour of a multi- objective optimization algorithm in traffic light control systems. The proposed local
search is integrated into the iteration process of the evolutionary algorithm to speed up the convergence rate. In this local search strategy, a potential direction is selected before starting the searching process by classifying the population into different sub- populations and hierarchical fronts. The reference solutions which are used to create neighbor solutions are selected in the same sub-populations and the first two fronts.
The created neighbor is likely to dominate the original solution. Thus, the chance to immediately find out a superior solution from the first search in this proposed local algorithm is increased. Hence, anytime behaviour of the evolutionary algorithm would be enhanced.
Fuzzy distance can be used as an indicator to decide which model should be used to evaluate the goodness of a candidate solution in the optimization process.
It is difficult to obtain a surrogate model with very high accuracy due to the lack of available data. If only the surrogate model is used to estimate the fitness value of candidate solutions, the evolutionary search will likely converge to a false optimum.
Therefore, the surrogate model is used together with a traffic simulator in an effective way to predict the goodness of candidate solutions. A critical question is what which solutions should be evaluated by the traffic simulator and which solutions might be estimated by the surrogate model. A fitness evaluation scheme was introduced to solve this question. In this scheme, a candidate solution should be estimated by the surrogate model if this solution is close to the dataset and the surrogate model is reliable. If a solution is close to the dataset, the searching area surrounding that solution is well studied, as a result, the approximation error should be small. Furthermore, if the estimation error of the surrogate model is high, it should not be chosen to evaluate the solution.
The closeness of a candidate solution and the dataset is measured by the minimum fuzzy distance between that solution and the dataset which is defined as the smallest fuzzy distance between the solution and all the samples in the dataset. If the minimum fuzzy distance between a solution and the database is smaller than a threshold and the surrogate model is reliable, that solution will be estimated by the surrogate model.
Otherwise, its fitness value will be evaluated by the traffic simulator.