A study on evolutionary computing based web service selection techniques

Abstract

Many service providers are offering their business functionality as web services. The problem of web service selection is a complex and time-consuming activity. Among other techniques, a significant work has been reported on the use of evolutionary computing based algorithms in determining optimal web service for a task. A rigorous review of the state-of-the-art for efficient selection of web services using evolutionary computing based algorithms published over the last decade is presented. The existing works on web service selection using various evolutionary approaches with a discussion on algorithmic variations, their effect on selection, quality of service parameters used, contributions, limitations and research gaps of these works are explored.

Appendices

Appendix 1

A number of different fitness functions are presented in the different research works to improve the performance of GA for efficient selection of WSs. A summary of various fitness functions used to improve selection of WSs in different ways is shown in Table 7. The fitness functions listed in Table 7 can be grouped into four categories. First category (Cat1) includes fitness functions at serial number 1, 3, 5, 6, 12 and 14. Second category of fitness function (Cat2) incorporate fitness functions defined at serial number 4, 9, 11, 15 and 17. Category 3 (Cat3) encompasses the fitness functions at serial number 2, 7, 10, 13 and fitness functions at serial number 8, 16 are in category 4 (Cat4). The fitness functions in Cat1 are similar on treating the QoS parameters as either increasing type or decreasing type. The increasing parameters are kept in numerator and decreasing parameters are kept in denominator to maximize the fitness function score. On the contrary, the Cat2 fitness functions do not treat the QoS parameters differently. But, they are similar to Cat1 in imposing penalty by reducing the fitness values of individual web services involved in the composition. Fitness functions of Cat3 involve the computation of fitness values by including factors apart from QoS values such as current iteration, number of competing services and/or path information, etc. For fitness functions of Cat4, the fitness values for favorable solutions are given incentives and unfavorable solutions are penalized. This ensures the proper distance between two solutions (favorable/unfavorable).

Table 7 Summary of fitness functions

Appendix 2: Miscellaneous works on WS selection

In this section, the miscellaneous works (eighteen papers out of total reviewed papers) on GA based WS Selection are discussed, which uses other algorithms along with GA to improve the performance of selection mechanism. The algorithms compared with simple GA and important outcomes of these works are summarized in Table 8.

Table 8 Summary of comparison between GA and other suggested techniques

In Ai and Tang (2008b), the repair based GA that uses minimum-conflict hill-climbing (MCHC) as repairing operator is introduced. However, MCHC operator takes extra time to repair, but improves efficiency of WS Selection. Further, a work in Jin et al. (2008) has presented a new genetic based ant algorithm (GBAA) approach to solve WS Selection problem by first generating sub-optimal solutions using GA and then using a sub-optimal solution to initialize the pheromone in max–min ant system (MMAS) (Jin et al. 2008; Yang et al. 2019). In majority of work the WS Selection problem is modeled as single objective optimization problem. In papers (Yao and Chen 2009; Claro et al. 2005), the WSC problem is modeled as MOOP and a modified non-dominated sorting based GA (NSGA-II) is employed. The NSGA-II algorithm ranks and sorts each competing individual based on non-domination level. In addition to NSGA-II algorithm, a new crowded comparison operator is applied in Yao and Chen (2009). The crowded comparison operator creates a new pool of offspring and to calculate the crowded distance of each member in the pool. This results in increased population fitness and improved quality of services are selected.

Apart from QoS parameters, two parameters - transactional properties (Ding et al. 2015b, a), Quality of Experience (QoE) (Peng et al. 2009), can act as parameter for optimal selection and composition of WSs. Using the transactional properties and transaction composition rules, a GA based transactional QoS aware service selection algorithm, known as TQoS (Ding et al. 2015b) and TGA (Ding et al. 2015a) can also be employed. In Peng et al. (2009), customer expectation and environment is used as one of the parameters for acceptability of composition. In this paper (Peng et al. 2009), a new parameter called as Quality of Experience (QoE) ameliorates GA by combining learning capability of SA with GA. Furthermore, the work also deals with the problem of pre-mature convergences of GA. The single and multiuser WS Selection problem can be envisaged as an optimization problem.

In few of the work (Arockiam and Sasikala Devi 2012; Palanikkumar and Gnana 2012; Zhao et al. 2015a, b; Jatoth et al. 2018), the hybrid solution based on GA and other algorithms is proposed to solve WS Selection problem. It is reported that a hybrid solution using GA and particle swarm optimization (PSO) (Palanikkumar and Gnana 2012), GA with ACO (Zhao et al. 2015a, b), GA with EDA (Jatoth et al. 2018) and also by using a hybrid of SA and GA (Arockiam and Sasikala Devi 2012) performs better than pure GA for WS Selection. In hybrid solutions, the advantages of all participating algorithms is combined which covers disadvantages/limitations of all participating algorithms. This helps in enhancing the capability of GA and improves the performance of WS Selection algorithm. On the contrary, hybrid solutions are complex and need to be designed carefully to leverage full advantage of all participating algorithms.

Performance of GA for selection of desired WS largely depends on initial population generated (Beran et al. 2012). Population with higher fitness leads to best quality optimal solution at the end of GA iterations. Therefore, the multiple composition plans generated as a result of the random walk with exhaustive search act as a useful starting point for GA based WS Selection and the blackboard algorithm based WS Selection. The blackboard methodology uses two concepts useful for efficient selection of WS. First, a knowledge base to guide the search in a systematic manner (Beran et al. 2012). Second, a decision tree based on cost estimation of visited paths (Beran et al. 2012). Similarly, the initial WSC plan can be generated automatically with the help of Genetic Programming (GP) due to its benefit of using a tree structure to represent individuals (Rodríguez-Mier et al. 2010; Yu et al. 2013; Xiao et al. 2012; Ma et al. 2015; da Silva et al. 2016). Similarly, the overall quality of WS Selection can be improved by first following Advanced A-Fully Polynomial Time Approximation Scheme to obtain pareto optimal solution (Zhu et al. 2018). The GA is followed to find optimal solution for all the users

In this section up till now, all the WS Selection algorithms solve the service selection problem by considering end-to-end QoS requirements. This enforces the selection algorithm to check whether end-to-end QoS are being satisfied by combining QoS of all the participating services. A different way to solve the WS Selection problem is proposed in Liu et al. (2015) by decomposing end-to-end QoS requirements using cultural algorithm with GA (known as CGA). After decomposing end-to-end QoS, local search techniques can be used to find WSs satisfying local constraints