Introduction:
Workflow refers to a range of activities that have been executed with an aim of creating a specific function where the jobs would generally be reliable on each other. For instance, information transfer that could take place between these jobs leads to the development of a workflow that looks forward to employing resources for different workflow activities by optimizing the utilization of cloud resources (Naik et al.2021). There are certain single objective workflow scheduling systems that are capable of combining linearly a range of objectives in order to achieve a multi-objective solution while it would be incapable of modeling efficiently a real-world problem for situation where there involves a dynamic environment (Xia et al.2022). Cloud computing keeps on revolutionizing the utilization and management of computational resources, providing flexible and scalable solutions for different computational operations. One of the critical areas associated with cloud computing remains workflow scheduling in which tasks get to be assigned to the computational resources in such a way they ends up optimizing certain objectives like reduction of run time, energy consumption or costs. On the other hand, dynamic workflow scheduling is essential challenging because of the dynamically changing feature of the cloud environment (Ismayilov and Topcuoglu, 2018).
The concept of multi-objective evolutionary algorithm:
Multi-objective evolutionary algorithms can be defined as the paradigms associated with evolutionary computing such as evolutionary strategies and generic algorithms applied for resolving problems demanding optimization of two or more significantly self-contradictory objectives, without resorting to the minimization of objectives into a specific objective employing a weighted sum (Zhu et al.2015). The key idea behind evolutionary algorithms get derived from the selection and variation principles that are considered as the basis of Darwinian evolution theory. With the operations of an evolutionary algorithm on a particular population, every individual that would represent a candidate solution to a provided optimization problem. Furthermore, multi-objective evolutionary algorithms involve a class of optimization tools employed for resolving problems including a range of conflicting objectives whereas the algorithms are created aiming at finding a range of solutions that present a trade-off among various objectives instead of a specific optimal solution (Sardaraz and Tahir, 2020). In a number of practical problems, there exist different objectives that require being simultaneously optimized while these objectives seem to contradict with each other. What it means is that enriching a specific objective could end up resulting in a degradation of another. The conventional methods of sole-objective optimization are incapable of dealing with these scenarios efficiently since they look for finding one best potential solution (Zhu et al.2015). Be that as it may, multi-objective evolutionary algorithms utilize principles that have been created taking inspiration from the biological evolution theory of Darwin which is all about looking for solutions in a multidimensional area of potential solutions (Shukla et al.2016). All these algorithms holds a populace of candidate solutions which are called as solutions or individuals whereas over different generations, they keep on evolving this populace for finding a diverse range of solutions representing trade-offs between the conflicting objectives (Naik et al.2021).
The key components and concepts associated with multi-objective evolutionary algorithms:
Population: Population in MOEAs is nothing but a set of individual solutions representing a point in a multi-dimensional search area (Sardaraz and Tahir, 2020).
Objective tasks: Objective tasks are metrics employed for measuring the solution qualities in terms of various objectives. In multi-objective evolutionary algorithms, these objective tasks generally represent different objective operations that require being optimized simultaneously (Ramezani et al.2015).
Dominance: A solution ends of dominating another solution if the former one is better than the later one in at least an objective and would not be worse in any other objective. However, assignment help USA and this relation of dominance is employed for ranking and comparing different solutions (Ismayilov and Topcuoglu, 2018).
Pareto dominance: A solutions end up Pareto dominating another solution if the former solution dominates the later one and there remains at least one objective in terms of which the former objective strictly gets better than the later one and the solutions that would not be dominated by other solutions would be signified as non-dominant or Pareto optimal solutions (Ismayilov and Topcuoglu, 2020).
Solution: This is the process of selecting solutions from an existing population for creating the foundation for the generation of new solutions within the next generation. The progress remains biased towards choosing solutions that would be representative, diverse and non-dominant in nature (Mohammadzadeh and Masdari, 2023).
Mutation and crossover: Just like the conventional evolutionary algorithms, multi-objective evolutionary algorithms at times employs mutation and recombination or crossover operators for generating advanced solutions by connecting attributes from chosen parent solutions (Ramezani et al.2015).
Maintenance of diversity: The aim of multi-objective evolutionary algorithms is to maintain a different types of solutions within a population for covering multiple regions associated with the Pareto Front, representing the trade-offs among different objectives resulting in allowing to ensure that the algorithms do not prematurely converge to a particular region within the solution space (Ismayilov and Topcuoglu, 2018).
Stopping criteria and convergence: Multi-objective evolutionary algorithms look forward to converging towards the actual Pareto Front while the stopping criteria for the algorithms usually are dependent on multiple generation, convergence metrics or computational resources (Mohammadzadeh and Masdari, 2023).
The integration of neural network and multi-objective evolutionary algorithm to achieve dynamic scheduling in cloud:
Workflow scheduling is all about an approach to assign tasks to cloud environmental resources for ensuring complete execution of these tasks efficiently which requires defining the sequence of executions of the tasks and the resources allocated. Dynamic environment in cloud computing refers to the accessibility of resources that keeps on changing frequently because of different parameters like resource failure, fluctuating demands and maintenance operations (Mohammadzadeh and Masdari, 2023). Therefore, scheduling algorithms are supposed to adapt to the changes in real-time. Scheduling of work includes simultaneous optimization of a range of conflicting algorithms like execution time, energy consumption and cost reduction whereas neural network includes a special type of artificial intelligence that can mimic the function and structure of human brains with being capable of learning from patterns and making predictions and decisions based on such patterns. Evolutionary algorithms are nothing but a subset of artificial intelligence supported by neural networks and applying different mechanisms including selection, mutation and crossover for evolving solutions for optimizing next generation problems (Ramezani et al.2015).
Integrating network networks into multi-objective evolutionary algorithms is capable of embracing the capabilities of both these techniques. On the one hand, neural networks are capable of predicting and adapting to dynamic environments by identifying patterns from existing data. On the other hand, evolutionary algorithms are capable of searching effectively though larger solution space for the sake of finding the most optimal solutions (Ismayilov and Topcuoglu, 2018). Tentatively, neural networks have been combined with multi-objective evolutionary algorithms for increasing the size of the population as well as accelerating convergence. One of the critical ideas in this process, is known as neural network-based multi-objective evolutionary algorithm which is all about calculating all the individuals within network networks rather than directly evaluating them, however, the distinguish exists in the combination of the number of times and set of trainings that a neural network gets trained (Sardaraz and Tahir, 2020). In a particular algorithm following the execution of a standard NSGA-II for different generations, neural network gets trained for estimating more candidates of which some would be chosen to be evaluated further. It is a well-performing combination in the process of optimizing the dynamic area of aperture as well as the lifecycle of Touschek. When there is no change in a former objective, the next objective ends up increasing around 10% than the standard multi-objective evolutionary algorithm within similar duration (Zhu et al.2015).
It is noticed that the neural network-based multi-objective evolutionary algorithms perform efficiently in the convergence velocity while involving a drawback of totally depending on the calculated indicators for choosing patterns as the training of neural networks starts exists still in them (Xia et al.2022). After the consideration of the strict constraints in the optimization, it would be easier to estimate the feasible candidates as infeasible due to the small size of a training set which is considered as a potential concern for such algorithms (Shukla et al.2016). However, for dealing with this challenge, there exist a penalty function that can progressively be stricter over the executed generations in fitness functions for fulfilling gradually the constraints. Furthermore, neural networks gets incorporated as the operators which would not only include multiple candidates to be evaluated further like the operators associated with crossover and mutation do, but also ends up screening a larger number of internally estimated individuals (Naik et al.2021). The effectiveness of these operations remains difference under the changes in a penalty and hence the number of candidates that source from those operators to be evaluated further with being redistributed dynamically (Xia et al.2022). Such an algorithm gets termed as the dynamically employed neural network-based multi-objective evolutionary algorithm. Furthermore, there would be a proposal of accessibility algorithm as a new approach to handle the preferences in NSGA-II which would not be dependent on any additional reference points that would mutually be set in another algorithms for leading the non-dominated fronts for approaching them (Shukla et al.2016).
Conclusion:
A neural network-based multi-objective evolutionary algorithm for dynamic workflow scheduling in cloud computing environment presents an effective approach to deal with both the dynamism as well as complexities associated with the advanced cloud platform. By embracing the capability of predictive analytics and pattern identification associated with neural networks and the enrichment of the capacity of evolutionary algorithm, this system becomes capable of serving as adaptable and efficient solutions to multi-objective workflow scheduling problems.
References:
Ismayilov, G. and Topcuoglu, H.R., 2020. Neural network based multi-objective evolutionary algorithm for dynamic workflow scheduling in cloud computing. Future Generation computer systems, 102, pp.307-322.
Ismayilov, G. and Topcuoglu, H.R., 2018, December. Dynamic multi-objective workflow scheduling for cloud computing based on evolutionary algorithms. In 2018 IEEE/ACM International Conference on Utility and Cloud Computing Companion (UCC Companion) (pp. 103-108). IEEE.
Mohammadzadeh, A. and Masdari, M., 2023. Scientific workflow scheduling in multi-cloud computing using a hybrid multi-objective optimization algorithm. Journal of Ambient Intelligence and Humanized Computing, 14(4), pp.3509-3529.
Naik, K.J., Chandra, S. and Agarwal, P., 2021. Dynamic workflow scheduling in the cloud using a neural network-based multi-objective evolutionary algorithm. International Journal of Communication Networks and Distributed Systems, 27(4), pp.424-451.
Ramezani, F., Lu, J., Taheri, J. and Hussain, F.K., 2015. Evolutionary algorithm-based multi-objective task scheduling optimization model in cloud environments. World Wide Web, 18, pp.1737-1757.
Sardaraz, M. and Tahir, M., 2020. A parallel multi-objective genetic algorithm for scheduling scientific workflows in cloud computing. International Journal of Distributed Sensor Networks, 16(8), p.1550147720949142.
Shukla, S., Gupta, A.K., Saxena, S. and Kumar, S., 2016. An evolutionary study of multi-objective workflow scheduling in cloud computing. International Journal of Computer Applications, 133(14), pp.14-18.
Xia, X., Qiu, H., Xu, X. and Zhang, Y., 2022. Multi-objective workflow scheduling based on genetic algorithm in cloud environment. Information Sciences, 606, pp.38-59.
Zhu, Z., Zhang, G., Li, M. and Liu, X., 2015. Evolutionary multi-objective workflow scheduling in cloud. IEEE Transactions on parallel and distributed Systems, 27(5), pp.1344-1357.