A High Performance Random Forest Machine Learning Algorithm for QoS- Based Web Services Composition

doi:10.21203/rs.3.rs-3326023/v1

Download PDF

Research Article

A High Performance Random Forest Machine Learning Algorithm for QoS- Based Web Services Composition

https://doi.org/10.21203/rs.3.rs-3326023/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Service-based architecture becomes a key software interface for complex applications which can be dynamically and flexibly compounded by integrating existing Web services with standard protocols from various providers. New web services can adversely affect the quality of service and customer satisfaction through their rapid introduction into a competitive business environment. Hence, it remains a constant problem in state of the art research how to collect, aggregate and employ data on individual quality of service (QoS) components to obtain the optimum quality of service of a composite service that responds to customer needs. This study proposes a Random Forest (RF) Algorithm for high-performance machine learning web services composition (WSC) problem. In this experiment, the validation is done using value iteration, iterative policy evaluation, and policy iteration algorithm. In this experiment, we illustrate WSC difficulties how to solve and demanding an entire sequence of 1000,000 Web services can be measured using an Intel Core i7 device with a 32GB RAM, requiring the selection of 10,000 services from the current system and less than 130 seconds. Besides, only seven individual web services pose a real WSC problem that needs processing power of less than 0.03 seconds.

Web services

Machine Learning

QoS

Web services consist of independent URI packages that can be advertised, located, and accessed by XML-based messages and transferred via Internet protocol [1]. The new model of service-oriented computing (SOC) ensures that companies and organizations will collaborate with standard web services in an unparalleled manner [2]. Web services that fulfill the functionality of the requestors must be able to find, dynamically, and continuously changing from a wide range of service providers depending upon their quality of service (QoS), so that they can improve the dynamic and rapid composition of services within this paradigm [3, 4]. As web services become more broad and complex, the efficiency of SOC is a critical challenge for QoS [5, 6]. As a service-oriented architecture (SOA) the new generation model for developing agile Internet distributors has been recognized. For applications, web services can be defined, consisting of more coarse grain services (i.e. composite services). The aim is to find the best service to create a new service and to deliver the best QoS under the constraints of users and service designers [7].

The web service can usually not satisfy the need for specific business activity in the real world but can serve diverse needs by creating different web services. Therefore, there is a possible silver bullet for businesses and several researchers have been interested [8–014]. In order to create such systems, web services include the necessary basic unit representing particular activities or features. When services are exposed and made accessible on the web for other web services and tools, they can build more dynamic frameworks and interactions and provide new value-added services. Services must be viewed as modular, composite, and adaptive units to efficiently support the highly complex design of these structures so that other services are quickly replaced, transformed, or even composed.

The key point was to find Web services before choosing and writing web services, while Ngan and Kanagasabai [15] provided a thorough analysis of highlighting state-of-the-art methods, semantic web service discovery, key semantical formalities, and performance measurement criteria as well as tests. The composition and selection of web services contribute to a problem of optimization which has been studied in both the SOC research field for grid and business processes applications. The definition and storing of QoS data and the customer specifications [16] are two main issues when using QoS for service discovery. Zeng et al. [2, 17] implemented web service composition QoS-based middleware. The QoS output was not derived from workflow patterns, but the service composition was divided by directed acyclic graphs into implementation processes. They have used integer programming to solve the objective function of any running path and to integrate solutions into all running paths to support service demands for optimal services.

With the number of exclusive compound choices and the possibility that service choices will overlap, the expense of this approach will inevitably be increased exponentially. In order to correct cycle processes and negotiate in order to deal with the situation when it is not possible to come up with a feasible solution Ardagna and Pernici [18] expand the ILP model. This means that the ILP solution is only suited for minor issues as the complexity of the algorithm increases exponentially as the problem is increasing. In order to help service demands numerically analyze services, Huang et al. [19] applied a variety of parameters for the decision-making of weighted sums of model and the integer programming approach. Wu et al. [20] enhanced the current service composition method by specifying a QoS requirement with its selection algorithm based on QoS reference vectors. There are two algorithms to perform an experiment (exhaustive search and ILP) and two types of experimental examples are compared. One is for the most detailed selection of resources and the other chases to balance selection costs with performance.

To define their discrepancies on service options, The authors [21] suggested a fuzzy-scale linear programming method to help the service customer choose the most relevant services with respect to requirements and preferences. In order to establish QoS-assured end-to-end communication channels over administratively different domains, Xiao and Boutaba [22] provide an autonomous service delivery framework. Hence it provides the model and provides algorithms to overcome the problems of domain construction and adaptation as the classic k-multi-constrained Optimum path (MCOP). The QWSC problem was modeled by Yu et al. [55] in two ways: the combinatorial model and the graphical model. A multi-dimensional 0–1 knapsack problem is based on the combinatorial model. The model graph defines the problem as an MCOP problem. The solution ensures that all services chosen follow QoS criteria and that no essential route can be identified before the selection of a service.

Tsesmetzis et al. [23] concentrated on the viewpoint of the service provider receiving multiple simultaneous services requests that illustrate various QoS features. In order to detect the services to optimize income for the provider, subject to maximum bandwidth constraints, the 'selective multiple-choice knapsacking problem. Ma et al. (24), which knew of the ameliorated Dijkstra and self-fitting algorithm to the best way to integrate candidate web services, proposed a dynamically driven and weighted acyclic service correlation model. Its function is to take account of dependence between services and identify the correlations between services. Jiang et al. [25] have analyzed and modeled the nature of QoS Aware's automated service structure on a new graphic problem: how an optimally controlled single source acyclic diagram (SSOD) from a managed graph can be found? They formally submitted their model and then their overall computing rules or QoS in order to deal with this new SSOD issue. The problem with the composition of the service is formalized in a complex search method for a stochastic discreet event system's optimal plan [26].

In order to reduce high computational complexity, authors have created a reliable service selection algorithm with Markov's better decision-making process. In comparison, the LMD approach [27] does not need a pre-determined threshold nor does the distribution of the provider require accurate parameters. LMD is the first technique in the domain that can be adapted for itself because it makes the decision without human intervention based on complex information obtained during the awaiting of the process. Xu et al. [6] use an efficiency approach by using a dynamic programming method to construct a polynomial-time method for the shortest composition sequence. Several other approaches for this problem are also suggested, for example, branch and bound [29], analytic hierarchy process [28], clouding computing [30], grey relational analysis [32], clustering [31], principal component analysis [34], multi-state bounded sets technique [33], skyline analysis [35], systematic search algorithm like Dijkstra [37], and statistical regression [36].

2.1 Contributions of This Paper.

The automated WSC aims to identify a sequence of web services to meet a number of predefined QoS restrictions. The Random Forest (RF) algorithm is one of the most robust mathematical methods for machine learning that we can use for problems where the sequence of actions to optimize the overall output function is to be calculated. We, therefore, suggest an RF model for solving the problem of WSC in this paper. We have experimented with the three most studied algorithms in order to demonstrate the reliability of our model: policy iteration, value iteration, and iterative policy evaluation. While all the three methods were provides numerous best solutions, a minimal number of iterations were needed to inclusion for ideal solutions by the policy iteration algorithm. We also contrasted the three algorithms with the Q-learning and sarsa algorithms, showing that one or two magnitude orders and longer time are needed to solve the problems of composition with the same complexity.

The following paper is organized as follows. The basics of the Composite system given in section 3, Section 4 given QoS criteria. The RF system is given in Section 5 and the three algorithms we have tested are introduced. Section 6 presents our WSC solution RF model. Section 7 explains and presents the most important findings of the experimental installation. Section 7.3 introduces sarsa and Q-learning comparative studies. Section 8 Conclusions and future studies.

A composite service can be built and deployed on various platforms with the following different methods. Based on the business process, which involves a collection of abstractive services with a set of components, the objective is that each abstract service should be combined in an optimal way. The functionality of an abstract service is described, not the Web Service's (WS) details [38]. The features of a web service are normally carried out through the operations it provides. The abstract service functional characteristics include the inputs, outputs, and the type of classification that is an ontological technique.

Algorithm 1 (Abstract service):

A tuple is defined as WS and noted (P, OR, L, Q) in which, P is the abstract service's;
OR is the kind of activity necessary as regards ontological concepts;
I = {I ₁, . I = I ,I_m} is a set of inputs of necessary operations where m is the input number. I_i = (IN, IT), in which I = 1,... m, IN, input name, and IT, ontologically, input type;
O = {O ₁,.,.., O_n} which is set of outputs of the operations required in which n is the number of outputs. O_i = (OS, OR), where OS is the output name and OR the ontology definition output form.

The Abstract Service is not inherently a single service implementation, which refers to a particular functionality. In order to find some services with a matching algorithm (called concrete or component services), this description is used to pick one that meets the constraints and maximizes the QoS aggregation feature. There have been several suggested matching methods [15, 39–41].

Algorithm 2 (Selection plan):

The selection plan SP is a sequence of pairs (ws_i, op_i), where SP = {(ws₁, op₁), ., (ws_n, op_n) }. The service operation op_i contributes the necessary function of the abstract service ws_i for each pair (ws_i, op_i) in SP, here I = 1, ., n. n.

The selecting plan depicts a mix of procedures for an abstract procedure offered by concrete services. Therefore, each abstract service has been alloted a composite service that has a working flow description language as described in the interaction (e.g. [7]). The following two issues must be concluded:

Obtain a composite service's QoS based on the component's QoS;
Obtain the collection of unique services for any workflow abstract service, maximizing the composite service QoS.

Provided p abstract services and q specific services, pq are possible combinations for each abstract service. The search is discreet and there is a problem with combinatorial optimization. The vast search area makes it necessary and RS complexity, an exhaustive search algorithm is not sufficient.

During the QWSC service selection process the QoS standards are employed to distinguish web services that provide similar functions. Each web-based service activity will be connected with a collection of QoS functions (non-functional properties) according to ISO 8402 [43] that can properly represent the needs of users and affect user gratification from certain aspects, such as run-time, availability, reliability, reputation, and cost, commonly employed in a number of references [2, 18, 38, 42]. These characteristics refer to web components and to Web composites.

Execution time T(ET): The estimated time lag among the very short in time of beginning a service operation and the moment of obtaining the result. Time to execute is typically in seconds measured.

Reliability Rel(ET): Trustworthiness is a appraise of the trustworthiness of service invocation, that is the likelihood of an operation being accurately answered within the period anticipated.

Availability A(ET): The possibility of usability of the business process. The number in the range [0, 1] is available.

Reputation Rep(ET): The trust of the service operation op is calculated. The ratio of the number of service calls that meet the agreed QoS to the total number of service calls shall be specified. The reputation is a [0, 1] number.

Price P(ET): The charge (cost) a service applicant pays for the service process op to the service provider. The price in dollars ($) is assessed.

A QoS vector is employed to calculate the QoS metrics of service, according to the value and calculation of the service quality parameters.

An RF is a tree-based ensemble, as the name suggests, where each tree is based on a collection of random variables [44]. More formally, we assume an unknown joint distribution PX_XY for a n-dimensional random vector Y= (Y_1,……..,Y_p)^T representing the real-valued input or predictor variables and a random variable X representing the real-valued answer (X,Y). In order to predict X, the objective is to find a prediction function f(Y). A loss function L(X,f(Y)) specifies the prediction function and defines it to minimize the estimated value of the loss,

E_XY(L(X,f(Y))) (1)

Where subscribers imply expectations for the joint distribution of Y and X. L(X,f(Y)) is intuitively a measure of how close f(Y) is to X, the value penalizing f(Y) values that are long from X. Squared error loss L(X,f(Y)) =(X − f(Y))² and null-one loss for classification are common choices of L:

L(X,f(Y)) = I(X ≠ f(Y)) =$\left\{\begin{array}{c}0 if X=f\left(Y\right)\\ 1 otherwise\end{array}\right.$ (2)

It turns out the conditional assumption is to minimize E_XY(L(X,f(Y))) for squared error loss,

f(y) = E(X|Y = y) (3)

Often known as the function of regression. When the set of potential Y values is indicated by y, in the classification situation, there is a reduction of zero-one loss to EX_Y(L(X,f(Y)). In that case.

f(y) = argmax P(X = x|Y = y), (4)

The law of Byes is otherwise known.

The groups build f for so-called "basic learners" h₁(y),...,h_J(y) collections and the basic learners to give the "assembly predictor" f(y). The basic learners are summed in regression,

f(y) = 1/J∑_j=1hj(y), (5)

f(y) is the class most commonly foreseen ("voting") when in classification.

f(y) = argmaxy∑_j=1I(x = h_j(y)). (6)

The j^th basis learner in Random Forests is a tree known as h_j(Y,Θ_j), where Θ_j is a random variable set, while Θ_j is independent for j = 1,...,J. Though very general, the concept of a random forest is almost always used in the manner defined, In the particular manner mentioned, they are almost invariably carried out. It is necessary to have a basic knowledge of the type of trees used as basic learners in order to understand the algorithm of the Random Forest.

5.1 Dynamic Programming Algorithms for RFs

Dynamic programming can be used to address the probabilities of state change (6). Next, we have three effective dynamic programming algorithms for the resolution of finite-state RFs. The first is the iterative review of policies. The second is the iteration algorithm of the policy value. This algorithm repeatedly calculates the current policy value functionality and then uses the current value function to update the policy. The third is the value function iteration as an iteratory update of the value function, which is seen in Algorithm 3 employing (6). Moreover, iteration and value iteration are common algorithms for RF-resolution and the algorithm that is stronger [45, 46] is currently not widely accepted.

In this section, we have set the RF model employed to describe and resolve the composition problem of web services using dynamic programming algorithms. We start with a thorough explanation of the WSC problem. The features, input information, and output information of each Web service can be classified into categories. Provided that C unusual groups of a single web service are used, the issue with WSC is to find the sequence of the length C of each web service (w1, w2, …., wC), in that way wi € Wi, for i = 1,2,….,C, where Wi constitutes a collection of all class i web services. We expect, therefore, to provide a valid web composite service in each of the current classes. Furthermore, we assume that all the web services available have previously been categorized and that classes 𝑊1 ≺ 𝑊2 ≺ ⋅⋅⋅ ≺ 𝑊𝐶 were predetermined. However, we believe that historically all web services were categorized as WC. To ensure that the selected web services operate properly, 𝑊𝑖 ≺ 𝑊𝑗 means that the Web Service from set Wi must be performed before a Web Service is set from Wj. The right functions of input and output information are important to the correct process. The production of wi must therefore be completely consistent with wj. Now our model is ready to be described. We describe the problem of the composition of a web service as RF (S, A, P, γ, R), where the set of states is S, the set of acts A, the feature of state probability function P, µ is a factor that discounts µ ∈ [0, 1] and R is the function of rewards. Next are specified elements S, A, P, and R.

6.1 States: the set of states is S. Established a 𝐶 classes with WSC issue, 𝑆 comprises all compositions of a maximum length of 𝐶. For 𝐶=1 thus, 𝑆 = {⟨𝑤1⟩}, for 𝑤1 ∈ 𝑊1. A length 𝑙=1 composition with a single Web service; we will however relax the importance of the term “composition” and consider it a length of composition 𝑙=1.

6.2 Actions: The set of all actions A. In the case of state s, a set of available acts is referred to as A(s); hence A= {A(s)}s ∈S is referred to as A(s). One action is to choose a Web service in the present composition that is included. Thus the present length of the composition is l = i, the current action kit will consist of all options for choosing a class c = i + 1 web service.

Thus, we call that 𝐴 = {𝐴(𝑠𝑙=0), 𝐴(𝑠𝑙=1), 𝐴(𝑠𝑙=2),. . ., 𝐴(𝑠𝑙=𝐶−1)}, Where A(sl = i) is read as a series of acts from a state representing a length l = i composition. Note that A(sl = 0) refers to the set of actions that are accessible from the length l = 0 composition, referring to where no Web services have yet been chosen. For instance, if the present state is the composition of (w1,w2) then all the options of choosing a Web service of class c = 3 are provided for the A(sl = 2). Otherwise stated, a position where the c = 1 have chosen Web services and c = 2 web services have already been chosen, and now require to choose a c = 3 class of Web service.

6.3 Transition Probabilities: For all states of S and P(s'|s,a) are state transition probabilities, s'∈S and accomplishes of a ∈ A, that are presently accessible from s and s'. Remember that the likelihood of moving from the state s=(w1) testimony 1 ̈ to the state s'=(w1,w2). In the meantime, the possibility of moving from the likely condition s=(w1) and a state s' = (w1,w2,w3) is 0. Otherwise, we can only switch from a length composition state l = i to another length l = i + 1 composition state.

6.4 Reward Function: 𝑅(𝑠’ | 𝑠, 𝑎) is the compensation obtained when activity 𝑎 is taken and the surroundings create a modulation from 𝑠 to 𝑠’. Our model’s rewards feature is determined by employing three QoS attributes, originally proposed in [47]. Availability, execution, and throughput time are the QoS employed,

𝑅 (𝑠) = $\frac{a{v}^{s}-a{v}^{min}}{a{v}^{max}-a{v}^{min}}$ - $\frac{{time}^{s}-{time}^{min}}{{time}^{max}-{time}^{min}}$ + $\frac{t{r}^{s}-t{r}^{min}}{t{r}^{max}-t{r}^{min}}$ (7)

where time^𝑠, 𝑎v^s, 𝑡𝑟^𝑠 are the average execution time, availability, and throughput values applied to the constitution interpreted by state 𝑠 for the last Web service. time^min, 𝑎v^min, 𝑡𝑟^min and time^max, 𝑎v^max, and 𝑡𝑟^max are the minimum and maximum values for all the Web services.

The findings of our experimental comparison are discussed in this section using two examples, one actual and another artificial. The tests we presented in this section have been conducted using the Intel Core i7 3.6 GHz Processor, on Windows 10 pro, 64 bits of operating system and 32 GB of RAM, and using value-iteration, iterative policy iteration, and running policy iteration algorithms.

7.1 Real Scenario: In this study, the first experiment comprises two classes of web services. One class concerns weather services that can be used in a region to get current temperatures. The other class includes web services, for instance, Celsius to Fahrenheit, which can be used to transfer heat from one metric unit to the next. We considered three separate web services in the weather services group. (a) GlobalWeather Web service, available at. https://www.ibm.com/docs/en/api-connect/saas?topic=api-example-wsdl-file-globalweatherwsdl (b) National Oceanic and Atmospheric Administration (NOAA) Web service, available at https://www.noaa.gov/weather (c) Weather channel Web service, available at https://www.weather.gov/documentation/services-web-api. We found four web services in the category of metric units conversion services.

(a) A basic Web service calculator like the one available

http://www.dneonline.com/calculator.asmx .

Since 𝐶 = 5 ∗ (𝐹 − 32)/9, (8)

For temperature conversion, we can employ multiplication, subtraction, and division.

(b) Convert Temperature Web service, available at

http://www.webservicex.net/ConvertTemperature.asmx .

http://webservices.daehosting.com/services/TemperatureConversions.wso .

(d) Temp Convert Web service, available at

http://www.w3schools.com/webservices/tempconvert.asmx .

All seven web services were given for the QoS attribute values by means of a Java program designed to obtain the following formulation attribute values:

Availability = 𝐶_𝑆/𝐶_𝑇, (9)

If C_S is the number of active Web service calls and _CT the total number of calls,

Execution time = 𝑇/𝐶_𝑇, (10)

where 𝑇 is the total execution time for all the 𝐶𝑇 calls,

Throughput = 𝐶_𝑆/𝑇, (11)

with 𝐶_𝑇 = 50.

In order to obtain representative QoS value for web services, we have made several tests, several days in different periods of the day. The values were obtained and the average QS values were then calculated for each parameter and measurement. Once we have gained the data of the attributes of QoS, all three dynamic programming algorithms have been used to learn the best composite Web service. There are twelve possible compositions of seven Web services belonging to two separate groups. All these choices are shown in the graph shown in Fig. 1.

The real-life scenario graph shows the layer of any form of web services. Per node represents a single web service in this graph. Node S is the condition that has not been chosen yet for any of the Web services. Node G reflects the state of the complete composition of the web service. A route from S to G includes the generation of a valid web composite service.

Figure 2 explains the outcomes of the actual Web services scenario. All of the three algorithms found the solution for the composition of web services very easily and were the winner in less than 0.01 seconds.

7.2 Artificial Scenario.

We simulated information on three QoS attributes: execution time, availability, and throughput in this second assumption test all three dynamic programming algorithms. This study, built up to 1,000,000 web services, divided into 1000 groups of hypothetical Web services. We have presumed any i-class web service can access all the i + 1-class web services. The 1000 layers or individual web services are included in each layer. As in the first case, node S is the graph's initial state and reflects a situation in which no web services have yet been chosen. When a true composition is completed, node G is reached. The available Web services are defined by nodes between S and G. A S-to-G route offers a potential composite web service. Figures 3, 4 and 5 show findings for this second series of experiments for µ = 0.7, µ = 0.8 and µ = 0.9 respectively.

The graph shows every layer of 100 web services of the same class. So with a valid web service composition number of nodes to be chosen at 10000, we are solving a problem with 100 x 10,000 = 1000,000 web services. The learning curves show that the time taken to solve the RF problem increases with the increase in the number of nodes. Again, the best solutions were found in all three algorithms, but in less time policy iteration was found. In µ = 0.8 and µ = 0.9, need needs less than 130 seconds to obtain the optimum composition using the iterative policy assessment and value iterations and less than one hundred seconds when it was politically iterated, the best achievements of algorithms were achieved.

7.3 Comparison with Sarsa and Q-Learning

Reinforcement learning algorithms were suggested for a resolution of the compositional problem of web services in some relevant works [47–049]. We compare Q-learning and sarsa time with policy iteration, value iteration, and iterative policy evaluation. This Section compares the learning time needed.

7.3.1 Sarsa. Sarsa is an on-policy algorithm to track temporal difference that continually estimates the status value Qπ for the conduct policy − and at the same time moves π to greed with respect to Qπ, respectively. The sarsa algorithm as taken from [50] is seen in the Algorithm. If the strategy is such that in any state all acts are done in an endless way, each state is repeatedly visited infinitely, and it becomes greedy for the present action-value function, then the algorithm converges to the Q* [52] in decaying α.

7.3.2 Q-Learning. Q-learning [53] is a non-political time gap management algorithm that explicitly approximates, independently from the policy being implemented, the optimum action-value function. It's one of the most common reinforcement learning algorithms. A Q-learning algorithm is modified as from [50]. If the limit values are indefinitely always modified for all state-action pairs with the decay of α, then with probability 1 [51, 53] the algorithm converges to Q*.

7.3.3 Learning Time Analysis. In order to solve the scenario problem described in the experimental field, we have implemented Q-learning and Sarsa algorithms. We can clearly see from this graph that two orders of magnitude and more time were needed for sarsa and Q-learning to find the optimum composition. Figure 6 presents the high definition measure with a logarithmic scale. In addition, we performed experiments with a second artificially constructed scenario, each with 3 layers of 20 Web services. Again, reinforcement learning approaches took a lot more time than dynamic programming algorithms. In addition, the reinforcement learning method in some experiments failed, being in suboptimal compositions, to find the optimal solution. Dynamic methods of programming converge more rapidly than methods of reinforcement learning simply because dynamic methods of programming update each state value with every iteration. Only the value of the states that happen to visit is modified by reinforcement learning techniques, giving its exploration strategy, that is, epsilon greedy.

In addition, it should be noted that the processing of QoS information may be carried out at specific time intervals by a dedicated module of such a system with regard to the implementation of an automatic Web service composition system. If we have gathered this data, which is central to the evaluation of the reward function, as reinforcement learning methods do, there is no need to explore the state space of web services. We simply to operate a dynamically programmable algorithm to estimate the web service value function and to determine the optimum web services composition.

In this study, a Random Forest Algorithm to resolve the composition of Web services is proposed. Three programming algorithms were used to demonstrate the efficiency of our approach. These include the value iteration, iterative policy evaluation, and policy iteration. Artificially generated data and a collection of real data with seven publicly accessible Web services were performed for experimentation. Our test results show that political iteration is the best way to determine an optimal policy by the minimum quantity of iterations needed. The optimal strategy displays the series of integrated Web services that make up a composite Web service with the highest QoS assessment.

Future research into this subject needs to discuss the composition of the real Web services with more nodes. The design of complex fee functions to deal with a growing number of QoS variables is another important topic that requires further research.

Conflict of interest: The authors declares that there is no conflict of interest.

Human and animals rights: No experiments with either humans or animals have been conducted by us.

Funding sources: No specific grant was given to this research by funding organisations in the public, private, or not-for-profit sectors.

Informed consent: Not applicable.

Web Services Architecture Requirements Working Group, 2004. Available at http://www.w3.org/TR/wsa-reqs
Zeng, L.Z., Boualem, B., Anne, H.H., et al.: ‘Qos-aware middle ware for web services composition’, IEEE Trans. Softw. Eng., 2004, 30, (5), pp. 311–327
Hu, C.H., Chen, X.H., Liang, X.M.: ‘Dynamic services selection algorithm in web services composition supporting cross-enterprises collaboration’, J. Central South Univ. Technol., 2009, 2, pp. 269–274
Liu, Y.T., Ngu, A.H.H., Zeng, L.Z.: ‘Qos computation and policing in dynamic web service selection’. WWW 2004, New York, NY, USA, 2004, pp. 66–73
Kondratyeva, O., Kushik, N., Cavalli, A., et al.: ‘Evaluating web service quality using finite state models’. 13th Int. Conf. on Quality Software, Nanjing, China, 2013, pp. 95–102
Xu, B., Luo, S., Yan, Y.X., et al.: ‘Towards efficiency of QoS-driven semantic web service composition for large-scale service-oriented systems’, Serv. Oriented Comput. Appl., 2012, 6, pp. 1–13
Canfora, G., Penta, M.D., Esposito, R., et al.: ‘A lightweight approach for QoS-aware service composition’. Proc. 2nd Int. Conf. on Service Oriented Computing (ICSOC’04), New York, NY, USA, 2004, pp. 36–47
Bertino, E., Chu, W.C.C.: ‘Guest editorial: special section on service-oriented distributed computing systems’, IEEE Trans. Serv. Comput., 2009, 2, (3), pp. 245–246
Jaeger, M.C., Rojec-Goldmann, G., Muhl, G.: ‘Qos aggregation for web service composition using workflow patterns’. EDOC 2004, Monterey, CA, USA, 2004, pp. 149–159
Li, M., Zhu, D., Deng, T., et al.: ‘GOS: a global optimal selection strategies for QoS-aware web services composition’, Serv. Oriented Comput. Appl., 2013, 7, pp. 181–197
Sherchan, W., Loke, S.W., Krishnaswamy, S.: ‘Explanation-aware service selection: rationale and reputation’, Serv. Oriented Comput. Appl., 2008, 2, pp. 203–218
Yu, Q., Bouguettaya, A.: ‘Guest editorial: special section on query models and efficient selection of web services’, IEEE Trans. Serv. Comput., 2010, 3, (3), pp. 161–162
Yu, T., Lin, K.J.: ‘A broker-based framework for QoS-aware web service composition’. IEEE Int. Conf. on e-Technology, e-Commerce and e-Service, Hong Kong, 2005, pp. 22–29
Zhang, L.J.: ‘Editorial: modern services engineering’, IEEE Trans. Serv. Comput., 2009, 2, (4), p. 276
Ngan, L.D., Kanagasabai, R.: ‘Semantic web service discovery: state-of-the-art and research challenges’, Pers Ubiquit. Comput., 2013, 17, pp. 1741–1752
Xu, Z.Q., Martin, P., Powley, W., et al.: ‘Reputation enhanced QoS-based web service discovery’. Int. Conf. on Web Services, Salt Lake City, UT, USA, 2007, pp. 249–256
Zeng, L.Z., Benatallah, B., Dumas, M., et al.: ‘Quality driven web services composition’. 12th Int. Conf. on World Wide Web, Budapest, Hungary, 2003
Ardagna, D., Pernici, B.: ‘Adaptive service composition in flexible processes’, IEEE Trans. Softw. Eng., 2007, 33, (6), pp. 369–384
Huang, A.F.M., Lan, C.-W., Yang, S.J.H.: ‘An optimal QoS-based web service selection scheme’, Inf. Sci., 2009, 179, pp. 3309–3322
Wu, B.Y., Chi, C.H., Xu, S.J., et al.: ‘Qos requirement generation and algorithm selection for composite service based on reference vector’, J. Comput. Sci. Technol., 2009, 24, (2), pp. 357–372
Wang, P., Chao, K.M., Lo, C.C.: ‘On optimal decision for QoS-aware composite service selection’, Expert Syst. Appl., 2010, 37, pp. 440–449
Xiao, J., Boutaba, R.: ‘Qos-aware service composition and adaptation in autonomic communication’, IEEE J. Select. Areas Commun., 2005, 23, (12), pp. 2344–2360
Tsesmetzis, D., Roussaki, I., Sykas, E.: ‘Qos-aware service evaluation and selection’, Eur. J. Oper. Res., 2008, 191, pp. 1101–1112
Ma, S., Chang, W., Cui, X.X.: ‘Service-correlation aware service selection for composite service based on the improved Dijkstra algorithm’. 2010 Int. Conf. on E-Business and E-Government, Guangzhou, China, 2010, pp. 2330–2334
Jiang,W., Wu, T., Hu, S.L., et al.: ‘Qos-aware automatic service composition: A graph view’, J. Comput. Sci. Technol., 2011, 26, (5), pp. 837–853
Fan, X.Q., Fang, X.W., Ding, Z.J.: ‘Indeterminacy-aware service selection for reliable service composition’, Front. Comput. Sci. China, 2011, 5, (1), pp. 26–36
Di, X.F., Fan, Y.S, Shen, Y.M.: ‘Local martingale difference approach for service selection with dynamic QoS’, Comput. Math. Appl., 2011, 61, pp. 2638–2646
Wu, C., Chang, E.: ‘Intelligent web services selection based on AHP and Wiki’. IEEE/WIC/ACM Int. Conf. on Web Intelligence, Silicon Valley, CA, USA, 2007, pp. 767–770
Liu, M., Wang, M.R., Shen, W.M., et al.: ‘A quality of service (QoS)-aware execution plan selection approach for a service composition process’, Future Gener. Comput. Syst., 2012, 28, pp. 1080–1089
Hashizume, K., Rosado, D.G., Fernandez-Medina, E., et al.: ‘An analysis of security issues for cloud computing’, J. Internet Serv. Appl., 2013, 4, p. 5
Xia, Y., Chen, P., Bao, L., et al.: ‘A QoS-aware web service selection algorithm based on clustering’. IEEE Int. Conf. on Web Services, Washington, DC, USA, 2011, pp. 428–435
Xiao, X.C., Wang, X.Q., Fu, K.Y., et al.: ‘Grey relational analysis on factors of the quality of web service’, Phys. Procedia., 2012, 33, pp. 1992–1998
Elfawal-Mansour, H., Mansour, A., Dillon, T.: ‘Composite web QoS with workflow conditional pathways using bounded sets’, Serv. Oriented Comput. Appl., 2013, 7, pp. 101–116
Kang, G.S., Liu, J.X., Tang, M.D., et al.: ‘Web service selection algorithm based on principal component analysis’, J. Electron. (China), 2013, 30, (2), pp. 204–212
Wu, J., Chen, L., Yu, Q., et al.: ‘Selecting skyline services for QoS-aware composition by upgrading MapReduce paradigm’, Cluster Comput., 2013, 16, pp. 693–706
Zaremba, M., Migdal, J., Hauswirt, M.: ‘Discovery of optimized web service configurations using a hybrid semantic and statistical approach’. IEEE Int. Conf. on Web Services, Los Angeles, CA, USA, 2009, pp. 149–156
Yan, Y., Chen, M.: ‘Anytime QoS-aware service composition over the GraphPlan’, Serv. Oriented Comput. Appl. Arch., 2015, 9, (1), pp. 1–19
Lin, C.F., Sheu, R.K., Chang, Y.S., et al.: ‘A relaxable service selection algorithm for QoS-based web service composition’, Inf. Softw. Technol., 2011, 53, pp. 1370–1381
Bai, L., Liu, M.: ‘Fuzzy sets and similarity relations for semantic web service matching’, Comput. Math. Appl., 2011, 61, pp. 2281–2286
Johnston, E., Kushmerick, N.: ‘Web service aggregation with string distance ensembles and active probe selection’, Inf. Fusion, 2008, 9, pp. 481–500
Segev, A., Toch, E.: ‘Context-based matching and ranking of web services for composition’, IEEE Trans. Serv. Comput., 2009, 2, (3), pp. 210–222
O’Sullivan, J., Edmond, D., Hofstede, A.T.: ‘What’s in a service?’, Distrib. Parallel Databases, 2002, 12, (2–3), pp. 117–133
ITU-T Rec. E.800: Terms and definitions related to quality of service and network performance including dependability, 2008
Cutler, A., Cutler, D. R., & Stevens, J. R. (2012). Random forests. In Ensemble machine learning (pp. 157-175). Springer, Boston, MA.
D. P. Bertsekas and J. N. Tsitsiklis, Neuro-Dynamic Programming, Athena Scientific, 1996.
M. L. Puterman, Markov Decision Processes: Discrete Stochastic Dynamic Programming, Wiley Series in Probability and Mathematical Statistics: Applied Probability and Statistics, Wiley-Interscience, 1994.
H.Wang, X. Zhouy, X. Zhou,W. Liu, andW. Li, “Adaptive and dynamic service composition using Q-learning,” in Proceedings of the 22nd International Conference on Tools with Artificial Intelligence (ICTAI ’10), pp. 145–152, Arras, France, October 2010.
V. Todica, M.-F. Vaida, and M. Cremene, “Formal verification in web services composition,” in Proceedings of the 18th IEEE International Conference on Automation, Quality and Testing, Robotics (AQTR ’12), pp. 195–200, May 2012.
L. Yu,W. Zhili, L.Meng,W. Jiang, and X.-S. Qiu, “Adaptive web services composition using Q-learning in cloud,” in Proceedings of the 9th IEEE World Congress on Services (SERVICES ’13), pp. 393–396, Santa Clara, Calif, USA, July 2013.
R. S. Sutton and A. G. Barto, Reinforcement Learning An Introduction,The MIT Press, Cambridge, Mass, USA, 1998.
D. P. Bertsekas and J. N. Tsitsiklis, Neuro-Dynamic Programming, Athena Scientific, 1996.
S. Singh, T. Jaakkola,M. L. Littman, and C. Szepesv´ari, “Convergence results for single-step on-policy reinforcement-learning algorithms,”Machine Learning, vol. 38,no. 3, pp. 287–308, 2000.
C. Watkins, Learning from delayed rewards [Ph.D. thesis], University of Cambridge, 1989.
T. Jaakkola, M. I. Jordan, and S. Singh, “On the convergence of stochastic iterative dynamic programming algorithms,” Neural Computation, vol. 6, pp. 1185–1201, 1994.
Yu, T., Zhang, Y., Lin, K.J.: ‘Efficient algorithms for web services selection with end-to-end QoS constraints’, ACM Trans. Web, 2007, 1, (1), pp. 1–26

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

A High Performance Random Forest Machine Learning Algorithm for QoS- Based Web Services Composition

Status:

Version 1

Abstract

Figures

1. Introduction

2. Related works

2.1 Contributions of This Paper.

3. Composite web service

4. QoS criteria

5. The Random Forest (RF) Algorithm

5.1 Dynamic Programming Algorithms for RFs

6. Web Service Composition (WSC) Model

7. Experimental Evaluation

7.2 Artificial Scenario.

7.3 Comparison with Sarsa and Q-Learning

8. Conclusions

Declarations

References

Additional Declarations

Status:

Version 1