AUTOMATED COMPOSITION OF WEB SERVICE WORKFLOW

A Novel QoS-Enabled Multi-criteria Cost Search Algorithm

Jaina Sangtani and Gursel Serpen

Electrical Engineering and Computer Science Department, University of Toledo, MS 308, Toledo, Ohio, U.S.A.

Keywords: Service Oriented Architecture, Automated Workflow Composition, Heuristic-based Search, Multi-criteria

Cost Search, Uniform Cost Search Algorithm.

Abstract: The introduction of software technology has dramatically increased the efficiency of completing tasks. Code

reusability provides efficiency within the software engineering discipline. With the tumultuous increase in

acceptance of service oriented architecture, and thus, a rise in the number of web services, skilled software

developers spend a lot of time composing web service workflows, rather than creating innovative and

efficient services. Hence, we put forward a technique of code reusability that utilizes heuristic based search

methods to automate service workflow composition by weighting quality of service criteria by relevance

and importance to the users. We implement a novel and heuristic-based graph creation and search algorithm

where the heuristic function value is calculated through the uniform cost search based on each of the quality

of service criteria specified by the user. Application of the proposed automated workflow composition

algorithm is illustrated with success on an industry-grade service-oriented architecture problem.

1 INTRODUCTION

With the advent of the computer and the internet,

enterprise employees have access to several

applications from various providers, to perform their

job functions. A single business process involves

numerous individuals, applications, and frequently it

extends beyond the company‟s boundaries into

partner and customer companies. Information

communication between traditional application

boundaries does not give insight into the entire

business process.

New methods of software development utilize

loosely-coupled code to develop integrated and

information-centric applications. Thus, application

boundaries are no longer restrictions to the business

process, as information can flow between these

applications, and the entire business process is

transparent to those involved. One such method of

development with loosely-coupled code is Service

Oriented Architecture (SOA) (Newcomer,

2005)(Rao, 2004).

The Sirena Innovation Report quotes several

experts in the SOA field asserting that SOA

adoption is tremendously growing. Wide acceptance

of the SOA notion only began in 2005, and today

“every sizable software vendor has stated its future

roadmap is going to be SOA related” (Schmelzer,

2005). Agarwal et al. (2005) state “… web services

have received much interest in industry due to their

potential in facilitating business to business or

enterprise application integration.” Rao and Su

(2004) claim that “… Nowadays, an increasing

amount of companies and organizations only

implement their core business and outsource other

application services over the Internet.” Thus, the

ability to efficiently and effectively select and

integrate inter-organizational and heterogeneous

services on the Web at runtime is an important step

towards the development of the Web service

applications. The number of services available over

the Web has been increasing dramatically during the

recent years, and one can expect to have a huge Web

service repository to be searched (Rao et al., 2004).

With an enormous rise in the adoption of SOA

and an equitable rise in the number of services

created and stored in web service repositories, it has

become difficult to effectively and efficiently select

and integrate services manually. In addition, the

selection and integration of services creates a

massive service search space, which is difficult to

explore manually, in order to find the correct

integration of services to deliver the appropriate

result to the users.

122

Sangtani J. and Serpen G. (2010).

AUTOMATED COMPOSITION OF WEB SERVICE WORKFLOW - A Novel QoS-Enabled Multi-Criteria Cost Search Algorithm.

In Proceedings of the Fifth International Conference on Evaluation of Novel Approaches to Software Engineering, pages 122-131

 SciTePress

Manual workflow composition is a challenging

task for larger scale problems since its manageability

for an error-free implementation and consistency

becomes almost impossible to achieve as it is the

typical case for any user-driven process (Kim,

2004). In order to compose the best solution, a

developer must consider each quality of service

(QoS) criterion and its importance to the user, and

track all of these criteria through the solution

process. Simple automation of the process through

scripting is not possible, as it requires intelligence to

develop a weighted cost value from heterogeneous

QoS criteria values, and discover the optimal

solution utilizing weighted costs.

The current manual method of SOA composition

involves extremely skilled experts, such as software

engineers, developers, and architects to create the

architecture and integrate services for even the

simplest applications or workflows (Hafner, 2009).

Expert skill is also required to create innovative and

competitive services. “If a competitor introduces a

new service, the service provider must offer a

similar or better service within days or weeks, to

avoid losing customers” (Agarwal, 2005). With the

pressures of the business and the demand for

software developers to create new services

efficiently and quickly, there is need for a more

efficient method to create the architecture and

compose workflows with services. The automated

creation of workflows in SOA will allow skilled

experts more time to concentrate on the innovative

aspects of creating a variety of performance-

enhanced and extremely reusable services, rather

than the manual, prolonged, and error-prone tasks of

composing the workflow.

The need for automated composition of service

workflows emerges mainly for the following three

reasons: 1) the immense size of the search space

associated with the composition problem; 2) the

human errors associated with composing service

workflows manually; and 3) the drive to reduce the

workload of software professionals so that they

concentrate on more creative aspects of software

development. Hence, in this paper, we present a new

algorithm with a novel heuristic for automated

workflow composition from a database of services.

In section 2, we discuss related works in the

literature, and explain the integration of heuristic

search and workflow composition. We describe the

proposed algorithm for this in detail in section 3; in

section 4 we portray the implementation and testing

of the algorithm; and finally, we conclude the paper

with section 5.

2 LITERATURE SURVEY

Services are developed to be reused by integrating

them with other services to create workflows or

applications. There has been a considerable amount

of work in this area. Currently, in the industry,

service workflows and applications are composed

manually by software experts – developers and

architects. The literature has a number of

propositions to automate this process with logical

workflow composition, semantic workflow

composition, abstract process model, and artificial

intelligence (AI) planning. Some of these works

have been implemented, but at the present time there

appears to be very little evidence of testing. A brief

discussion of each of the prominent approaches

ensues next.

2.1 Manual Workflow Composition

As described in Hafner (2009), the manual service

workflow architecture has five layers – applications

layer, web services composition or publication and

discovery layer, service description layer, XML

messaging layer, and transport layer. Each of these

layers needs to be manually composed and created.

Today‟s frameworks automate portions of this to

ease the pain for developers. For example, the .NET

framework automates the service description layer,

XML messaging layer, and the transport layer.

However, there are still tedious manual tasks that the

service developer is expected to complete, including

the composition of the services into workflows, and

their integration with applications.

2.2 Logic-based Service Composition

The rapid development of service-based system

approach uses alpha-logic and alpha-calculus in

composing automated service workflows. A

developer creates the service workflow using alpha-

logic; the workflow is first converted into alpha-

calculus, and then into executable format.

Savarimuthu claims that the alpha-logic notation

is simpler than programming languages (2005).

However, comprehension and creation of workflows

with the alpha-logic notation requires high level

mathematical knowledge and awareness of services.

The service workflow composition is not automatic.

A software expert‟s time and efforts are required to

develop and implement service workflows with this

method. The logic-based workflow composition

method is a step ahead of the manual method, but it

AUTOMATED COMPOSITION OF WEB SERVICE WORKFLOW - A Novel QoS-Enabled Multi-Criteria Cost Search

Algorithm

123

requires advanced mathematical background on the

part of the software professional.

2.3 Semantics-based Dynamic Service

Composition

The semantics-based dynamic service composition

approach creates an automated workflow for users

through the semantics of a user request (Fujii,

2005)(Qui, 2007). The model used by Fujii, presents

the user with a simple textbox through which they

must enter their final outcome in plain English for a

software development request. The model analyses

the narrative in the user‟s query, and converts it into

machine readable format in order to find relevant

services, and compose a service workflow.

This method focuses on the translation of the

user query to attain services requested by the user,

but it disregards optimal search criteria and

integration of services. In addition, this method has

exponential time complexity, and it adds additional

steps with the conversion from English narrative to

machine readable format and the dynamic workflow

composition.

Qui, et. Al (2007) implement a similar model

with a context-aware architecture. However, this

model also disregards optimal search criteria and

integration of services. Additionally, this model is

only tested on one example for applicability, but not

for scalability or accuracy.

2.4 Abstract Process Model

The EFlow and Polymorphic Process Model form a

static service composition model, where the abstract

process model is created by the developer (Rao,

2004). The abstract process model requires

definition of tasks and data types. The automation

only includes the selection and binding of services to

the tasks described in the process model. Similar to

the Semantics-Based Dynamic Composition method,

the Abstract Process Model focuses on service

selection, but not search and integration through the

graph. Service search and integration of services

still remain as manual processes.

2.5 AI Planning

Artificial Intelligence (AI) planning and associated

heuristic search transcends from an initial state to a

final state by selecting actions within a domain

which is typically modelled by a graph. The final

product of a planning algorithm is a sequence of

actions that achieves a desired effect (Russell, 2003).

In the case of service oriented workflow, actions will

be considered to map to services. Hence, each

service selection is considered as a logical step in a

workflow that changes the current state (starting

with the initial state) of the workflow bringing it

closer to the user defined business outcome or the

goal state. In Figure 1, which displays a service

graph, each edge in the graph is a service, and each

node is a state. The initial state is the knowledge

input by the user. Each state is a combination of the

previously known knowledge and the output of the

current service. The objective of workflow

composition using AI planning is starting from the

user-defined initial state, searching for the

combination of services that will deliver the goal

state within the service graph through an optimal

path.

Figure 1: Service Graph.

There are several published studies that “reason

about the composition of web services using goal-

oriented inferencing techniques from planning”

(Canny, 2003). However, the implementation in this

area is limited at best. Cheatham and Cox

(Cheatham, 2005), and Agarwal, et al. (2008)

implement AI planning algorithms for web service

composition. However, their work lacks

consideration for QoS criteria, and due to lack of a

computational complexity analysis, it is difficult to

assess the scalability properties of their approach.

2.6 Problem Statement

Automation of the service composition is still an

open problem and needed to bring the service

oriented architecture methodology to a state of wider

applicability and realization. Consequently, a new

heuristic-based search algorithm is proposed in this

paper to automate the web service composition. The

proposed algorithm will search the graph model of a

service domain to identify an ordered sequence of

services as a plan for the composition of services.

The algorithm is able to incorporate QoSs in service

Service 3

Service 4

Service 2

Service 1

Initial

State

Goal

State

New State

Generated by

Service 1

New State

Generated by

Service 2

ENASE 2010 - International Conference on Evaluation of Novel Approaches to Software Engineering

124

composition with a multi-criteria cost measure. The

functional and non-functional properties of services

are also taken into account. Functional properties

describe the input, output, and the activities

conducted by the service. Non-functional properties

are the costs borne as a result of execution of each

service. Non-functional properties, also known as

quality of service (QoS) of a service include aspects

of services such as time to execute, resources

required for execution, monetary cost of the service,

and several other properties that might be of interest

(Schuschel, 2004)(Peer, 2004).

3 SYSTEM ARCHITECTURE

The comprehensive solution for web service

workflow composition in a Service Oriented

Architecture (SOA) context entails the following

five steps:

1. Acquisition of customer requirements and

specifications

2. Service discovery, selection, and

compilation

3. Creation of SOA graph

4. Formulation of workflow path

5. Software Integration and Implementation

The acquired customer requirements and

specifications as well as a repository of services are

sent into the novel graph creation algorithm. A

workflow path is found within the graph utilizing a

heuristic based graph search algorithm. The output

of the algorithm is a sequence of services sent to a

third-party executable software that generates an

output value for the goal state specified in the user

requirements.

3.1 Acquisition of Customer

Requirements and Specifications

The initial state, goal state, and name, maximum

value, and weight values for the set of Quality of

Services (QoS) are acquired from the user. These

entities are supplied as inputs into the web service

workflow composition algorithm. It is conceived

that these values can be input by a non-technical

user through a simple form like a Graphical User

Interface (GUI).

3.2 Service Discovery, Selection,

and Compilation

Service discovery is mapping customer requirements

to services which are semantically and ontologically

defined in order to provide the appropriate results to

the customer. Service selection and compilation is

the obtaining these services from a variety of service

repositories and libraries and installing the services

within the customer‟s environment.

Service discovery, selection and compilation is

essential to the area of web service composition.

However, this topic is a separate field of research,

and thus, considered outside the scope of this paper.

There are various works of literature that focus on

this topic. For example, Fujii discusses semantical

representation of services through a model called

Component Service Model with Semantics

(COSMOS). COSMOS is implemented alongside

Component Runtime Environment (CoRE) for

service discovery and selection (Fujii, 2005).

Additionally Qiu, et al, discuss the implementation

of semantical web service selection with a context

aware architecture (Qiu, 2007)

3.3 SOA Graph Creation

The SOA graph, which is proposed herein as a novel

abstract model for the service workflow domain, is a

specialized representation of the services selected

based on the customer requirements and the

normalization process. In the implementation of the

heuristic-based search algorithm, the service names,

their Web Service Description Language (WSDL)

file URLs, and their associated QoS values are

stored in a database within the customer

environment, accessible to the search algorithm.

A state in the web service workflow composition

domain is defined as a fixed-size set of attribute-

value pairs. An attribute is any variables whose

value is known at the end of a service execution. A

state‟s attribute-value pairs are composed of the

previous state‟s attribute-value pairs along with the

outputs generated through the most recently

executed service.

Each state of the environment is represented as a

node into the graph. Services connect two nodes if

the input of a service is contained within the source

node state. The output of the service associated with

the outgoing directed edge from the source node

state composes the successive node state of the

service. For each state in the environment, beginning

with the initial state input by the customer, services

are iterated through to find one(s) with desirable (or

AUTOMATED COMPOSITION OF WEB SERVICE WORKFLOW - A Novel QoS-Enabled Multi-Criteria Cost Search

Algorithm

125

target or goal) input variable-value (attribute-value)

assignments.

Both the data type and the name of the

input/knowledge criteria variables must match. In

Figure 2, Node 1 contains the value for the

knowledge criteria variable x. A service is found

that has input x, and is connected to the graph at

Node 1. The service has an output y. Thus, the

knowledge criteria of the next node, Node 2, include

the knowledge criteria of Node 1 and the output of

the service (assignment to the y variable). The input

of the next service requires both x and y variables,

and the output is the value assignment for the z

variable. Thus, the knowledge criteria of the next

node, Node 3, are a combination of the knowledge

criteria of the previous node, Node 2, and the output

of the service (value assignment to the z variable).

Figure 2: Knowledge Criteria.

The graph creation algorithm expects as input the

initial state, the desired goal state, the name, and

weight and maximum allowed value of each QoS

criterion of importance to the user. All available

services that match the QoS criteria described by the

user are considered. Each QoS value of a service is

normalized as follows:





= 









(1)

where N

is the normalized QoS value, Q

is the

original QoS value of the service, M

is the

maximum QoS value as specified by the user, and j

depicts the current QoS criterion with j=1,2,...,q (q

is the number of QoS criteria input by the user). All

services with 



values greater than 1 will not be

considered for the algorithm, because their QoS

values are greater than the maximum allotted by the

user.

3.4 Formulation of Workflow Path

3.4.1 Heuristic Search Algorithm

The new heuristic-based search algorithm created

specifically for service workflow composition is

described in Figure 3. The proposed search

algorithm takes into account the functional and non-

functional properties of services, and selects services

in the graph based on the QoSs of the service. As the

QoSs that are important to the user can be difficult

to anticipate, the proposed algorithm allows for

multiple QoSs.

The heuristic-based search algorithm, starting

from a given initial state or node u, selects the

neighbour n

with i=1,2,…,b

, where b

is the

neighbour count of node u, with the smallest

evaluation function value and adds it to the path

while discarding the other neighbours. Next it

expands this node n

and calculates evaluation

function values for its neighbours while retaining the

neighbour with the smallest evaluation function

value as part of the plan or solution path. This

process continues until the goal node is reached. The

evaluation function f(



) is calculated for each

neighbour node n

through











= 









+ 









for  = 1,2,  , 



(2)

To calculate 









, the weighted sum of the QoS

criteria is added to 







of the parent node u:











= 







+  









 =1

for i=1,2,…,b

(3)

where j is the current QoS criterion, and j=1,2,...,q (q

is the number of QoS input by the user), N

is the

normalized value of the current QoS for the service

that connects nodes u and n, and α

is the weight of

the current QoS as input by the user. The 







value

of the initial node is defined as 0.

To calculate 









, the uniform cost search

algorithm is executed with node 



as the start node

once for each of the QoS criteria. The uniform cost

search algorithm returns the optimal path with the

least cost for each QoS criterion j, which is

designated as 

,

. In the process, the appropriate

QoS value is stored with each node along the path of

the best path. The heuristic function 









is then

calculated with the following formula:











=  





,



 =1

for i=1,2,…,b

(4)

Service Execution

Input: x, y and Output: z = 4

Service Execution

Input: x and Output: y = 2

Node 1 Knowledge Criteria:

x = 3

Node 2 Knowledge Criteria:

x = 3 and y = 2

Node 3 Knowledge Criteria:

x = 3, y=2, and z=4

ENASE 2010 - International Conference on Evaluation of Novel Approaches to Software Engineering

126

Multi-criteria Search Algorithm Pseudocode

Let the initial node be the start node of the path and set the

current node u to the initial node.

For the current node u, expand it by generating all its

neighbours n

with i=1,2,…,b

where b

is the neighbour

count

of node u.

For each n

For each QoS criterion j with j=1,2,...,q

Compute 

,

= 

,

/ 



Compute 

,

with the uniform cost search

Compute 









=  





,



=1

Compute 









= 







+  





,



=1

Compute 









= 









+ 









Select the n

with the smallest 









value

If the selected node n

is not the goal, then set the selected n

as the current node u and repeat.



,

is the normalized value of QoS criterion j for service i; 

,

is the original value of QoS criterion j for service I; 



is the

maximum value of QoS criterion j as input by user; 

,

is the

least cost value of QoS criterion j from service i to the goal

node as found by the uniform cost search algorithm; and 



the weight of the QoS criterion j as input by user.

Figure 3: Pseudocode of Proposed Algorithm.

3.4.2 The Uniform Cost Search Algorithm

Illustrated

The application of the uniform cost search algorithm

to compute the least cost path with respect to a

specific QoS criterion value j from a given node n to

the goal node is illustrated in Figure 4. The

neighbour node n is depicted by node 1 in the figure.

Once node 1 is expanded, both nodes 2 and 3 are

added to the fringe as neighbors of node 1. Service 1

connects node 1 and node 3, and has 



=0.1 as the

normalized value of the current QoS criterion. This

value is stored as the g value of node 3. On the other

hand, Service 2 connects node 1 and node 2, and has





=0.4, which is stored as the g value of node 2.

Since, the g value of node 3 is less than that of node

2, node 3 is chosen for expansion. Node 3 has only

one neighbour, node 4, which is added to the fringe.

Thus, the fringe now contains node 2 and node 4.

Note that nodes 3 and 4 are connected by Service 3,

whose 



value is 0.7. Accordingly, the g value of

node 4 is calculated as follows: g value of node 2

(0.1) is added to the 



value of Service 3 (0.7) to

yield 0.8. The g value of node 4 is greater than that

of node 2; hence, node 2 is expanded next noting

that the goal node is a neighbour of node 2. Service

4, connecting node 2 to the goal node, has a 



value

of 0.3. Now the g value of goal node becomes

0.4+0.3=0.7. Next, the goal node is expanded since

its g value is less than that of node 4.

Based on the g values, the path selected from

node 1 to the goal node is via node 2. The least cost

value, 



, of a specific QoS criterion where

j=1,2,...,q, from any particular node on the selected

path to the goal node is stored within the node. Node

2 has an 



value of 0.3, the value of Service 4,

which connects it to the goal node. Node 1 has an 



value of 0.7, the values of Service 2 and Service 4

added, which connect it to the goal node. Since,

node 3 and node 4 were not on the selected path,

they do not have 



values stored. Storing the 



value of each consecutive node ensures that when

the node is selected as a neighbour node n by the

heuristic search algorithm in the future, the 



value

of that node will be known, and the uniform cost

algorithm will not need to be executed. For example,

if node 2 is selected as the n, the uniform cost

algorithm is not executed to find 



of node 2, as it is

already known that 



of node 2 is 0.3.

3.4.3 Time and Space Complexity

The time and space complexity for the uniform cost

search algorithm is given by O(













), where b is

the branching factor, 



is the cost of the optimal

solution, and  is the least cost of an action (Russell,

2009). There are two loops encircling the uniform

cost search invocations. The outer loop iterates b

ave

times, where b

ave

is assumed to be the average

branching factor of the search tree associated with

the SOA graph. The inner loop executes q times,

where q is the number of QoS criteria. Assuming

that the maximum depth of the search tree

(composed from the problem domain graph with

realization and removal of loops) is m, there will be

m invocations of the overall algorithm. Hence, the

time complexity is given by

 × O(













)  O(













)

(5)

where m, q, and b are finite-valued positive integers.

Only one uniform cost fringe is maintained and

remains in memory at any given time. Every time

the loop is restarted, the previous uniform cost fringe

is discarded from memory. Hence, the worst case

complexity for space is given by

O( + 







/



)

(6)

where m is the maximum depth of the search tree,

which is kept in the memory for the heuristic search

algorithm, and O(







/



) is the space cost of the

uniform cost search algorithm.

AUTOMATED COMPOSITION OF WEB SERVICE WORKFLOW - A Novel QoS-Enabled Multi-Criteria Cost Search

Algorithm

127

3.5 Software Integration

and Implementation

The database of services, selected by a semantic

service selection system, or by a developer, along

with the algorithms for graph creation and search

algorithms, and the executable OW2 Orchestra

software are packaged and deployed within the

user‟s environment for a complete solution. The

user interface to the database is employed to input

user requirements into the system and to extract the

final output from the system is also included.

Figure 4: Uniform Cost Search Illustrated.

4 SIMULATION STUDY

The algorithm has been implemented as a web

service using C# and the .Net 3.5 framework with

the Visual Studio 2008 IDE from Microsoft

Dreamsparks (Microsoft Corporation, 2010). The

implementation includes a database for storage of

services, developed with SQL Server 2008, also

from Microsoft Dreamsparks (Microsoft

Corporation, 2010). The name and URL of the

service WSDLs are stored in a single table. A

second table is created for the QoS values as related

to the service.

A user interface (UI) to the web service that has

fields to allow for a user to input the variables and

fields of the initial state and the goal state was

created. Users can add as many QoS criteria as

needed by entering their names, weights, and the

maximum allowable value. The UI additionally

allows for new services to be entered into the

database.

The SOA graph creation and heuristic-based

search algorithms were tested on a 64-bit Windows

7 OS computer with an Intel cpu (core 2 duo 2.8

GHz processor), and 4 GB memory. In testing,

complexity of the problems was managed through

the following attributes:

a. Number of branches in the workflow (path)

b. Number of branches in the workflow

(conditional)

c. Similarities of services (the more similar,

the more complex)

d. Number of services required to complete the

workflow

Scalability of the algorithm implementation was

empirically assessed by changing the following

attributes:

a. Total number of services available

b. Number of input and output parameters per

service

Testing was accomplished on the following problem

domains that exhibit varying degrees of complexity:

1. UK National Health Service (booking

appointments, checkup, and prescription

tracking) (Srivatava, 2010)

2. PostFinance (ordering services, payments, and

transaction management) (Srivatava, 2010)

3. Harrods™ (online shopping, finding products,

and transaction) (Srivatava, 2010)

4. Vanco (trouble ticketing workflow) (Agarwal,

2005) (Srivatava, 2010)

5. Credit Card Request (client verification,

accepting/rejecting client, account processing)

(SAP, NetWeaver, 2010)

6. Parts maintenance (reporting issue, approval

process, analyze and fix defect, confirm work)

(SAP, NetWeaver, 2010)

7. Getting a standardized permit (application

processing, approval, invoice execution,

creating and sending permit) (SAP, NetWeaver,

2010)

8. Finding Energy Product (request, providing

data, determining, and offering a quote) (SAP,

NetWeaver, 2010)

9. Customer Quote Request (requirements, design,

providing a quote) (Jennings, 1996)

10. Travel Itinerary (creating itinerary for hotel and

flight, billing) (Srivastava, 2003)

These problems involve complex branching,

conditional and recursive branching, and a large

number of services and inputs. Some of these are

composition problems in the industry, whereas

others such as examples 4, 9, and 10 are benchmark

Service 5

𝑁

𝑗

= ? ?

Service 3

𝑁

𝑗

= 0.7

Service 4

𝑁

𝑗

= 0.3

Service 1

𝑁

𝑗

= 0.1

Service 2

𝑁

𝑗

= 0.4

Node 1

𝐿

𝑗

= 0.7

g = 0

Node 3

𝐿

𝑗

= ? ?

g = 0.1

Goal Node

𝐿

𝑗

= 0

g = 7.0

Node 4

𝐿

𝑗

= ? ?

g = 0.8

Node 2

𝐿

𝑗

= 0.3

g = 0.4

ENASE 2010 - International Conference on Evaluation of Novel Approaches to Software Engineering

128

problems used in the literature (Agarwal, 2005)

(Jennings, 1996) (Srivastava, 2003).

The proposed algorithm has been applied to all

ten problems listed above with success, however,

due to space limitations, a representative but non-

trival problem will be elaborated upon below.

Optimality and complexity results for all the

problems will be briefly presented. Accordingly, the

workflow composition by the proposed algorithm

for the Customer Quote Request problem (example

9) is described. This problem is the most complex in

terms of branching and it deals with the most

number of services as well.

4.1 Customer Quote Request Problem

The customer quote request problem requires a

network to be designed based on customer

requirements. The goal state of the problem is a

customer quote returned. The customer submits

requirements. If it is a veterinary customer, and they

do not need anything to be changed, the process is

terminated. If they do need changes, the

requirements are identified. If it is a portfolio item

that the company has, the service ID and the price of

the service are simply provided as a quote to the

customer. If it is not a portfolio item, a legal review

is conducted, requirements are analyzed, a survey

maybe conducted, a network designed, and a quote

provided. (Jennings, 1996)

For the example problem, the maximum QoS

values and weights are shown in Table 1. As inputs

into the algorithm, QoS names, maximum values,

weights, a customer requirement, customer details,

and the goal state of a received customer quote are

entered.

Table 1: Maximum QoS for Customer Quote Request

Problem.

QoS Criterion

Maximum Value

Weight

Duration (minutes)

320

Volume

Price (per costing)

Penalty

There are several services in the database for this

example. Table 2 displays those services in the

database, and their QoS values for duration, volume,

price, and penalty. Some services have variations

with different QoS values. For example, there are 3

“Provide Quote” services with different values for

duration, volume, price, and penalty. Some services

in the table provide logic for multiple services. For

example, the “Capture Customer Details - Capture

Customer Requirements” 1 and 2 services in Table 2

provide the logic for “Capture Customer Details”,

“Is Vet Customer”, Is Customer Okay” and “Capture

Customer Requirements”.

Figure 5 displays a part of the graph created from

the services above, while noting that due to space

limitations, the entire knowledge criteria for a given

node (state) could not be shown. The “Yes” and

“No” in some of the states in the graph is the

conditional branching. The algorithm deals with the

conditional branching by analyzing the value of the

output of requests, along with the data types and

names. For example, for the output of the “Is

Customer Okay” service, the “isCustomerOkay”

output variable is analyzed for its data type and

name, but also for its value, “Yes” or “No” to pick

the correct next service. “Identify Service

Requirements Profile” will expect the

“isCustomerOkay” variable to be assigned a value of

“Yes” for its input.

If, for instance, all the conditions have a “Yes”

output during the execution, the workflow output by

the algorithm is a sequence of the following

services:

1. Customer Details 1

2. Is Vet Customer

3. Is Customer Ok

4. Customer Requirements 2

5. Identify Service Requirements Profile

6. Is Portfolio Item

7. ID Service

8. Provide Quote 1

Branches are found for Customer Details,

Customer Requirements, and Provide quote.

Customer Details and Requirements 1 and 2, are

immediately not considered for having price values

(149.99 and 299.99, respectively) greater than the

maximum allowed value (35).

Both the “Customer Requirements” and the

“Provide Quote” service branches join back to the

same states in the graph. Hence, only their g values

are considered for selection, as all their h values are

the same. Additionally, the uniform cost search

algorithm is run only twice per QoS criterion – at the

initial state and at the customer requirements state.

For example, the uniform cost search algorithm is

run from the node “initial state” for each QoS

criterion. For most nodes, there are no branches in

this graph, and those nodes are selected for each

QoS criterion. Examples of these nodes are

“Customer Details 1”, “Is Vet Customer”, “Is

Customer OK”, etc. Thus, the least cost of each QoS

criterion from these nodes to the goal node is stored

within these nodes by the uniform cost search

AUTOMATED COMPOSITION OF WEB SERVICE WORKFLOW - A Novel QoS-Enabled Multi-Criteria Cost Search

Algorithm

129

Table 2: Services and QoS.

algorithm. The heuristic based search will also

select these nodes because of no branching. Before

these nodes are selected, their h values need to be

computed. Computing the h value does not require

the uniform cost search algorithm to be run. In fact,

it can be computed as a weighted sum of the stored

least cost values for each QoS criterion. Thus, the

uniform cost search algorithm is run only one other

time, when branching occurs, at the “Customer

Requirements” nodes.

In the final workflow path, Customer

Requirements 2 and Provide Quote 1 are both

selected for smaller g values and all the h values at

both branches are the same. In this example, after

branching, the states coincide later in the tree.

Hence, the h values have lesser necessity. This may

not always be the case, however.

CD = Customer Details, VC = Is Vet Customer?, CO = Is

Customer OK?, CR = Customer Requirements, SRP =

Service Requirements Profile, PI = Is Profile Item?, SID =

Service ID

Figure 5: Partial Graph of the Customer Quote Request

Problem.

4.2 Discussion

The time for running the algorithm for the Customer

Quote Request workflow with the “Yes” path is 0.3

minute, whereas running the algorithm on the “No”

path for “Is Portfolio Item”, and considering a path

via the “Customer Details and Requirements”

service is 0.97 minute. The “No” path has double the

depth (m) of the “Yes” path. Also, the number of

steps in the uniform cost search algorithm (𝐶

∗

/𝜖 )

Service Name

Duration

Volume

Price

Penalty

Capture Customer Details -

Capture Customer

Requirements 1

149.99

50.0

Capture Customer Details -

Capture Customer

Requirements 2

299.99

1.0

Capture Customer

Details 1

0.00

2.0

Capture Customer

Requirements 1

5.00

2.0

Capture Customer

Requirements 2

5.95

0.1

Is Vet Customer

0.00

0.0

Is Customer Okay

0.00

0.0

ID Service Requirements

Profile

0.00

0.0

Is Portfolio Item

0.00

0.0

ID Service

0.00

0.0

Provide Quote 1

5.00

0.1

Provide Quote 2

1.00

0.0

Provide Quote 3

9.95

0.0

Legal Review and Is Legal

0.00

0.0

Analyse Requirements and

Is Survey Required

0.00

0.0

Survey CPE 1

15.00

0.0

Survey CPE 2

59.50

0.0

Survey CPE 3

0.00

0.0

Design Network

10.00

0.0

Request Further Info 1

5.00

0.1

Request Further Info 2

1.00

0.0

Request Further Info 3

9.95

0.0

ID Service

Provide Quote

(not discussed in this paper)

Provide Quote

ID Service

Requirements

Profile

Customer Details 1

Legal Review

and Is Legal

Is Profile Item

ID Service

Requirements

Profile

Customer

Requirements 2

Customer

Requirements 1

Is Customer OK

Is Vet

Custom

CD,

VC = „Yes‟

CD,

CO = „Yes‟

CD, CR 1

CD, CR,

SRP

CD, CR,

SRP,

PI = „No‟

CD,

VC = „No‟

CD,

CO = „No‟

CD, CR,

SRP,

PI = „Yes‟

CD, CR,

SID

CD, CR, SID,

Quote 2

Initial

State

CD, CR 2

CD, CR, SID,

Quote 1

CD, CR,

SID, Quote 3

ENASE 2010 - International Conference on Evaluation of Novel Approaches to Software Engineering

130

will be a greater value. However, the rest of the

parameters (b and q) maintain the same values.

Table 3 displays of the results for all ten

problems. In all cases, optimal solutions were

computed by the proposed algorithm with very low

time and space costs.

Table 3: Complexity and Optimality Results for All

Problems.

Problem

Time

(minutes)

Space

(in KB)

Optimal

Solution

UK National Health

0.36

11,283

Yes

PostFinance

0.84

31, 748

Yes

Harrods

0.25

8,360

Yes

Vanco

0.73

31,732

Yes

Credit Card Request

0.46

9,293

Yes

Parts Maintenance

0.59

23,940

Yes

Standardized Permit

0.89

37,462

Yes

Energy Product

0.12

5,000

Yes

Customer Quote Req

0.97

41, 248

Yes

Travel Itinerary

0.93

45, 888

Yes

5 CONCLUSIONS

We have presented a novel heuristic-based search

algorithm for automated composition of a web

service workflow subject to multiple QoS criteria.

The algorithm has been successfully tested on a

number of complex real-life problems. Simulation

study and results indicate that the proposed

heuristic-based search algorithm can successfully

address challenging web service composition

problems within reasonable computational cost

bounds.

REFERENCES

Newcomer, E; Lornow, G (2005). Understanding SOA

with Web Services. Addison Wesley.

Rao, J., & Su, X. (2004). A Survey of Automated Web

Service Composition Methods. Proceedings of the 1st

international workshop on semantic web services and

web process composition, (pp. 43-54).

Schmelzer, R. (2005). The Explosive Growth of Service-

Oriented Architecture Adoption. Information

Technology for European Advancement. Sirena

Innovation Report.

Agarwal, V., Dasgupta, K., Karnik, N., Kumar, A.,

Kundu, A., Mittal, S., et al. (2005). A Service Creation

Environment Based on End to End Composition of

Web Services. Proc. of the Fourteenth World Wide

Web Conference. Chiba: ACM.

Kim, J., Spraragen, M., & Gil, Y. (2004). An Intelligent

Assistant for Interactive Workflow Composition.

Proceedings of the International Conference on

Intelligent User Interfaces. Madeira: ACM

Hafner, M., & Breu, R. (2009). Security Engineering for

Service-Oriented Architectures. Springer Berlin

Heidelberg. ISBN 978-3-540-79539-1

Savarimuthu, BTR. (2005). Agent-based integration of

web services with workflow management systems.

Information discussion paper series, 05, Retrieved

October 1, 2008.

Fujii K., Suda T. (2005). Semantics-based dynamic

service composition. IEEE journal on selected areas

of communication, 23, Retrieved April 27, 2008.

Qiu, L., Change, L., Lin, F., Shi, Z. (2007). Context

optimization of AI planning for semantic web service

composition. Service Oriented Computing and

Application, (pp. 117-128).

Canny, J. F., Edwards, D. D., Malik, J. M., & Thrun, S.

(2003). Artificial Intelligence: A Modern Approach.

Upper Saddle River: Pearson Education, Inc.

Cheatham, M., & Cox, M. T. (2005). AI Planning in

Portal-based Workflow Management Systems. Proc

International Conference on Integration of Knowledge

Intensive Multi-Agent Systems, (pp. 47-52).

Agarwal, V., Chafle, G., Mittal, S., Biplav, S. (2008).

Understanding approaches for web service

composition and execution. Proc. of the 1

Bangalore

annual Compute Conference.Bangalore: COMPUTE.

Schuschel, H., & Weske, M. (2004). Automated planning

in a service-oriented architecture. IEEE International

Workshops on Enabling Technologies: Infrastructure

for Collaborative Enterprises. 13, 75-80.

Peer, J.; Vokovic, M. (2004). Towards automatic web

service composition using AI planning techniques.

Proc. of the European Conf. on Web Services

(ECOWS’04). Springer-Verlag.

Russell, J. S., & Norvig, P. (2003). Artificial Intelligence:

Modern Approach. Prentice Hall.

Microsoft Corporation. (2010). Retrieved January 2010,

from Microsoft DreamSpark:

https://www.dreamspark.com/default.aspx.

Jennings, N. R., Faratin, P., Johnson, M. J., O‟Brien, P., &

Wiegand, M. E. (1996). Using intelligent agents to

manage business processes. Proc. of the First Int.

Conf. on the Practical Application of Intelligent

Agents and Multi-Agent Technology, (pp. 345-360).

Srivastava, B., & Koehler, J. (2003). Web Service

Composition - Current Solutions and Open Problems.

ICAPS2003 Workshop on Planning for Web Services.

Trento.

Srivastava, S. (2010) Sun‟s SOA & Solaris TM Enterprise

System. Sun Microsystems, Inc.

SAP NetWeaver (2010). Enterprise Use Cases. Retrieved

January 2010, from SAP Community Network:

http://www.sdn.sap.com/irj/sdn/index?rid=/webconten

t/uuid/a0cac44a-5820-2b10-96ae-c7494c38dec0.

AUTOMATED COMPOSITION OF WEB SERVICE WORKFLOW - A Novel QoS-Enabled Multi-Criteria Cost Search

Algorithm

131