**3. Review of evolutionary algorithms (EAs) for TSC**

This section reviews the previous studies in the literature that applied evolutionary algorithms (EAs) for traffic signal control and optimization. EAs are the most widely used metaheuristics optimization techniques across diverse fields of science and engineering. EAs are population-based random search techniques and are inspired by Darwin's theory of natural theory of evolution. The EAs contain a population of individuals, each symbolizing a search point in the feasible solution space exposed to a common learning process while proceeding among different generations. EAs begins with the initialization of random population, which are then subjected to selection, crossover, mutation through various generations so that offsprings generated evolve toward more favorable regions in the search space. At each generation, the fitness of the population is evaluated, and those with better fitness values are selected and recombined that have an increased probability of improved fitness. The program is iteratively repeated until it converges to the best (or near-optimal) solutions. The basic structure of EAs remains similar for all the algorithms under its family. **Figure 3** presents the sample structure of EAs and their working principle. The following passages provide a brief explanation of

**Figure 3.** *General flow depicting the search mechanism of EAs.*

#### *Metaheuristics for Traffic Control and Optimization: Current Challenges and Prospects DOI: http://dx.doi.org/10.5772/intechopen.99395*


**Table 1.**

*Summary of previous studies on traffic signal optimization using EAs.*

various EAs employed in the field of traffic signal optimization. **Table 1** presents a summary of previous studies that have applied EAs for traffic signal control and optimization.

#### **3.1 Genetic algorithm**

Genetic algorithm is the most widely used method for traffic light optimization. John Holland initially proposed the GA metaheuristic in 1975 [39]. GA search mechanism for finding the optimal solution of an objective function mimics the natural selection process of the evolutionary theory of nature, which supports the "survival of the fittest" concept. It is a population-based technique that involves the ranking of individual members of the population according to their fitness.

The search process of the optimal solution begins with the initialization of a random population of solutions. The offspring population is created by iteratively applying various genetic operators such as crossover, mutation, elitism, etc. until the stopping criteria are satisfied. In literature, many studies have demonstrated the robustness of GA for adaptive traffic signal control. For example, Foy et al. utilized GA for traffic light optimization, considering delay time minimization as the objective function [36]. The number of initial GA generations was varied over five GA traffic runs. The optimal fitness value was achieved for populations ranging between the 20th to 30th generations with an average vehicle waiting time of around 40 seconds. GA was noted to yield rational signal timing plans reducing the timing delay significantly compared to the existing traffic control scheme. In their study, Rahbari et al., studied that traffic control at the signalized intersection with GA could reduce the congestion [40]. Yang and Luo adopted a hybrid GA simulated annealing (GA-SA) for signal control optimization at isolated signalized intersections considering delay as the objective function [41]. Empirical results showed that GA produced a rational signal timing plan compared to fixed control scenarios. A study conducted by Mingwei et al. proposed the application of multi-objective for intelligent traffic management at an isolated signalized intersection for a case study in China [42]. The considered optimization objectives included; average vehicle delay, vehicular stops, and fuel consumption. It was found that the optimized signal timing plan from GA significantly improved the considered traffic performance measures.

In another study, Turki et al. proposed a multi-objective NSGA-II to optimize various measures of effectiveness (such as delay, stops, fuel consumption, and emissions) at isolated signalized intersections in the city of Dhahran, Saudi Arabia [35]. Study results were compared with Synchro traffic simulation and optimization tool, and the results for a typical intersection are shown in **Figures 4** and **5**. All the performance measures witnessed considerable improvement for the optimized signal timing plan obtained using an NSGA-II optimizer. **Figure 4 (a–d)** depicts the evolution of the four selected performance measures (delay, stops, fuel consumption, and emissions) against the number of iterations for three random initial populations. It may be noted that all the algorithms converged to their respective objective functions at approximately 70 to 100 generations. Comparing the random initial populations, population size 30 for all cases yielded the best results.

**Figure 5** shows the performance comparison of NSGA-II and Synchro signal control strategies for the selected measures of effectiveness (delay, stops, fuel consumption, and emissions). It may be noted from the Figure that the NSGA-II optimizer outperformed the Synchro results for all the performance measures.

Li et al. also investigated the applicability of NSGA-II for solving signal control optimization problems [34]. Average queue ratio and vehicle throughput were the objective functions. The algorithm's results were validated on a microscopic traffic *Metaheuristics for Traffic Control and Optimization: Current Challenges and Prospects DOI: http://dx.doi.org/10.5772/intechopen.99395*

*Evolution of different performance measures against NSGA-II iterations; (a) delay evolution, (b) number of vehicle stops evolution, (c) fuel consumption evolution, (d) emission evolution. Reprinted with permission from Ref. [35] copyright (2021), MDPI.*

#### **Figure 5.**

*Comparison of NSGA-II and synchro optimizers for various traffic performance measures. Reprinted with permission from Ref. [35] copyright (2021), MDPI.*

simulation tool, VISSIM. Kwak et al. developed a GA traffic optimizer to evaluate the influence of traffic light setting on vehicle fuel consumption and emissions [32]. Model results were compared with TRANSIM, a microscopic traffic simulator. It was observed that the proposed GA traffic optimizer could reduce exhaust emissions by approximately 20% and fuel consumption in the range of 8–20%. In another study, Kou et al. employed multi-criteria GA for optimizing the signal timing plan of signalized junctions and compared the results with the highway capacity manual (HCM) method [28]. The study considered several optimization objectives such as stops, delays, and emissions. A reasonable trade-off established an optimal Pareto front among different conflicting objectives. Study results demonstrated the superior performance of the proposed GA traffic control scheme compared to the HCM method in terms of all the optimization objectives. Guo et al. developed a model for area-wide intersection traffic control in the central business district (CBD) area of Nanjing, China [43]. Capacity ratio, turning movement delay, and travel time was the three chosen objective functions. Computational experiments results showed significant mobility improvement compared to existing conditions. Study results were also validated in PARAMICS traffic simulation tool. In their study, Dezani et al. have shown that simultaneous optimization of traffic lights via GA and vehicle routes could significantly reduce the vehicle travel time compared to optimization considering only routes [44]. In another study, Tan et al. proposed a new Decentralized Genetic Algorithm (DGA) for signal timing optimization of traffic networks under oversaturated traffic conditions [45]. Average vehicle delay was used as the performance metric to evalauate the performance of proposed algorithm. Simulation results showed that DGA could effectively optimize the traffic light setting and reduced the average network delay.

#### **3.2 Differential evolution (DE)**

Differential evolution is another population-based metaheuristic technique initially proposed by K.V. Pricein 1995 [46]. DE is characterized by its robustness, fast convergence to the objective function, and simplicity. Though the method has been successfully used for numerous applications across different disciplines, only a few studies have adopted DE for urban traffic control and management [25–29]. For example, in their recent study, Jamal et al. compared the performance of GA and DE for optimizing traffic lights at isolated signalized intersections in the city of Dhahran, Saudi Arabia [29]. Average delay time minimization was the objective function. The study concluded that both GA and DE could yield intelligent and rational signal timing plans, reducing the intersection average delay between 15 and 35%. DE was noted to converge to objective function faster than DA over several simulation runs. Similarly, in another study, Liu et al. proposed bacterial foraging optimization-based DE algorithm for optimizing delay at signalized intersections [37]. To improve convergence precision, DE was utilized for updating the bacteria position during the chemotaxis process. The proposed scheme yielded very promising results, reducing the intersection delay by over 28% compared to only 5% obtained by GA optimization. In their study, Korkmaz et al. suggested three different types of delay differential evolution-based delay estimation models (DEDEM), i.e., linear, quadratic, and exponential [47]. The researchers reported that all the proposed models effectively predicted the vehicle delay estimates in terms of relative errors between estimated and simulated values; however, quadratic DEDEM methods outperformed other models. Ceylan also approached the signal control optimization problem using the metaheuristic DE and Harmony-Search (HS) for network-wide traffic control and optimization [48]. Study results showed that DE algorithms yielded better results compared to HS.

In another research study, Yunrui et al. proposed multi-agent fuzzy logic control based on DE to optimize delay and queue length through a network of eleven intersections in the urban traffic context [31]. DE was used to decide and optimize the parameters of the fuzzy systems because it is easy to understand and implement.

*Metaheuristics for Traffic Control and Optimization: Current Challenges and Prospects DOI: http://dx.doi.org/10.5772/intechopen.99395*

Empirical results revealed that the proposed method could substantially improve the network performance measures such as average vehicle delay, traffic throughput, and queue length. In a recent study, Liu et al. have proposed an improved adaptive differential evolution (ADE)-based evolvable traffic signal control (EvoTSC) scheme for global optimization of different traffic performance measures on large scale urban transportation networks [49]. The proposed TSC method was compared with a conventional TSC scheme on two practical and three synthetic transportation networks with varying traffic flow demands and different physical scales. Comparison results indicated that the DE-based EvoTSC method significantly outperformed its counterpart under all the considered scenarios. Zhang et al. also applied an online intelligent urban traffic signal control approach using multiobjective DE for real-time traffic control and optimization [50]. Experimental results showed that the proposed approach provides a more robust configuration of traffic signal phases and has relatively better real-time performance than the traditional traffic control scheme.

#### **3.3 Genetic programming (GP)**

Genetic programming (GP) is another population-based metaheuristic technique that belongs to the family of evolutionary algorithms [51]. GP is an extension of GAs that allows for deep exploration of space on computer programs. GP starts with a population of random programs (candidate solutions) that are fit for applying evolutionary operators similar to genetic processes, thereby simulating the fundamental principles of Darwin's evolution theory [52]. GP follows an iterative process to breed the solutions to problems using the probabilistic selection procedure for the carryover of fittest solutions to the offerings by applying genetic operators such as crossover and mutation. In literature, not many studies have focused on applications of GP for traffic analysis and management in urban transport networks. Montana and Czerwinski used a hybrid GA with strongly typed GP (STGP) for intelligent control and optimization of evolving traffic signals on a small-scale transport network [53]. Numerical simulation results showed that the proposed hybrid STGP model could effectively improve network performance under varying traffic demands.

A study conducted by González also proposed the application of GP for solving signal control problems [54]. This study considered four different traffic scenarios with properties and traffic conditions in a previous study [55]. Study results were also validated using the microscopic traffic simulator tool SUMO. Findings showed that GP could provide competitive and robust results for all the tested scenarios. However, the highway/network scenario had a more pronounced performance improvement (having an improvement of 10.34%) than the isolated intersection scenario (with an improvement of 4.24%). In another study, Ricalde and Banzhaf adopted an improved GP with epigenetic modifications for traffic light scheduling and optimization under dynamic traffic conditions [56]. Extensive simulation analysis revealed that the proposed model improved the network performance compared to standard GP and other previously used methods. This study, however, did not use any real-world data for validation purposes. In another study, the authors used a similar GP approach with epigenetic modifications (EpiGP) to design and evolve traffic signals using real-time field traffic data [38]. Results indicated significant improvement in network performance compared to conventional methods, including standard GP, static, and actuated traffic control techniques.. Over 12% improvement in average delay was reported under high-density traffic conditions.
