Towards an Automated Optimization-as-a-Service Concept

Sascha Bosse

, Abdulrahman Nahhas

, Matthias Pohl

and Klaus Turowski

Very Large Business Applications Lab, Faculty of Computer Science Otto-von-Guericke University Magdeburg, Germany

Keywords: Cloud Computing, Optimization, as-a-Service, Business Analytics.

Abstract: Many organizations try to apply analytics in order to improve their business processes. More and more cloud

services are offered to support these efforts. However, the support of prescriptive analytics is weak. While

concepts for such an optimization-as-a-service exist, these require much expert knowledge in solution

methods. In this paper, a workflow for optimization-as-a-service is proposed that utilizes an optimization

knowledge base in which machine learning techniques are applied to automatically select and parametrize

suitable solution algorithms. This would allow consumers to use the service without expert knowledge while

reducing operational costs for providers.

1 INTRODUCTION

Data-intensive IT systems nowadays play an

important role in information systems, demonstrated

by the fact that organizations are in desperate need for

data scientists, big data analysts or comparable

experts (King and Magoulas 2015). In order to take

advantages of the opportunities of data-based

analysis, a process with three phases has been

identified (Evans and Lindner 2012):

 Descriptive analytics, which means characterizing

and understanding the past;

 Predictive analytics, in which it is desired to

estimate the likely future; and

 Prescriptive analytics, i.e. making decisions to

improve the future.

Recent developments in AI technology drive

analytics efforts today. While the former two phases

are strongly connected to the field of machine

learning, optimization is a crucial aspect especially in

the latter phase.

An organization applying analytics can build its

own capabilities or obtain them from external service

partners, while both approaches are cost-intensive

(Pohl et al. 2018). In this context, cloud computing

has revolutionized the way IT services are obtained in

the past years. While there are many cloud offerings

for descriptive and predictive analytics, there are

https://orcid.org/0000-0002-2490-363X

https://orcid.org/0000-0002-1019-3569

https://orcid.org/0000-0002-6241-7675

almost no IT service providers offering optimization-

as-a-service (OaaS).

Especially meta-heuristics such as genetic

algorithms are interesting in this context. These are

general, computing-intensive procedures based on

function comparisons that are applicable for a wide

range of real-world problems, for instance, virtual

machine placement (Müller et al. 2016), redundancy

allocation (Coit and Smith 1996), order sequencing

(Nahhas et al. 2017), or design planning (Lanza et al.

2015). Furthermore, especially population-based

meta-heuristics are by nature well suited for a high

degree of parallelization. Thus, the availability of

massive, distributed computing power makes the

cloud “the ideal environment for executing

metaheuristic optimization experiments” (Pimminger

et al. 2013).

For consumers, the concept of cloud computing

eliminates the problem of oversized systems for

computing-intensive tasks (Foster et al. 2008;

Pimminger et al. 2013) by the use of elastic, pay-per-

use self-services over the internet (Mell and Grance

2011). Thus, costs can be reduced for obtaining

optimization services for consumers, but also for

providers which can utilize economies of scale

(Marston et al. 2011).

Although concepts for such a service have been

introduced (e.g. in (Kurschl et al. 2014)), it remains

unclear how it can be effectively used by consumers

Bosse, S., Nahhas, A., Pohl, M. and Turowski, K.

Towards an Automated Optimization-as-a-Service Concept.

DOI: 10.5220/0007746303390343

In Proceedings of the 4th International Conference on Internet of Things, Big Data and Security (IoTBDS 2019), pages 339-343

ISBN: 978-989-758-369-8

339

with limited knowledge of problem formulations and

solution algorithms. On the other hand, providers are

bound to identify potential solutions to a problem

efficiently in order to reduce costs for both provider

and consumer. With a sufficient knowledge base of

optimization problems and the use of machine

learning techniques, the process of selecting a

suitable solution algorithm could be automatized.

Therefore, this short paper aims at presenting a

standard workflow for solving optimization problems

in the context of OaaS in order to discuss automation

potentials and future research opportunities.

2 RELATED WORK

Although the idea of combining meta-heuristics and

cloud computing is promising, only a few scientific

works deal with this topic. Most existing analytics

cloud service concepts lack of a prescriptive

component (cf. e.g. (Pohl et al. 2018)) and most

commercial offerings such as SAP Leonardo

do not

contain optimization services. Other such as IBM

Bluemix

or Google Optimization Tools

only

support mathematical optimization methods. In

contrary to meta-heuristics, these methods are often

not scalable to real-world (NP-hard) problems or

restrict the search space artificially (Coit and Smith

1996; Soltani 2014). Despite the lack of cloud

services, several frameworks for parallel, distributed

meta-heuristic optimization have been implemented

such as HeuristicLab

which could be utilized in an

OaaS context.

In (Pimminger et al. 2013), the authors investigate

the suitability of performing meta-heuristic

optimization in cloud scenarios. For that reason, large

scale experiments are conducted with different

deployment strategies. The authors conclude that

utilizing cloud resources for optimization massively

reduces costs for users.

Kurschl et al. follow this idea and provide a

requirements catalogue as well as a reference

architecture for OaaS (Kurschl et al. 2014). Although

technical questions such as multi-tenancy and

scalability are discussed, workflow-related topics are

not analyzed in detail. Consumers can access a cloud

platform to select and parametrize solution methods

for defined problems. Accounting is proposed to be

done on basis of the obtained computing power,

https://www.sap.com/germany/products/leonardo.html

https://www.ibm.com/us-en/marketplace/decision-

optimization-cloud

concepts for automation of the problem solution

workflow are not discussed.

3 A WORKFLOW FOR OaaS

In order to address the idea of solving optimizations

automatically without expert knowledge, a workflow

in context of the OaaS concept is presented in the

following.

When a potential user of an OaaS offering

encounters an optimization problem, this problem has

to be formalized by providing (Gill et al. 1993):

 A set of decision variables 



   



with the

respective domain,

 A set of  constraints 











   









 



with     , and

 An objective function 







   











   



 mapping decision variables to 

objective values.

The workflow of problem formulation and

solution in context of the OaaS concept is modeled as

a diagram in the Business Process Model and

Notation (BPMN) language and presented in Figure

In order to simplify the process of problem

formulation, basic problem classes should be

provided (Kurschl et al. 2014). However, a user

should also have the opportunity to formulate novel

optimization problems. Anyway, decision variables

should be characterized by defining a domain for each

variable, e.g. the set of real numbers or defined

discrete (or even binary) values. Based on the

decision variables, constraints can be defined by

entering symbolic functions such as simple bounds on

decision variables or linear combinations.

The first critical point of the workflow is the

definition of the objective function. Available

optimization frameworks often require a symbolic

objective function which simplifies the optimization

process, but may not always be available or

applicable. Therefore, implicit definitions of

objective functions should be supported additionally.

These can either be defined with a white-box or a

black-box approach.

In the white-box approach, the objective function

is modeled analytically, e.g. in a state-space model

such as Markov chains. For that reason, the

OaaSshould make use of a modeling engine which

https://developers.google.com/optimization/

https://dev.heuristiclab.com/trac.fcgi/

IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security

340

Figure 1: Workflow of an optimization service request.

allows users to define a variety of models for the

objective function. These models can be evaluated

numerically or by simulation. In the black-box

approach, the mapping from decision variables to

objective values is done by providing a high amount

of data which allows a supervised machine learning

algorithm to model the objective function.

After the problem has been formulated, (feasible)

solutions should be identified. In related work, it is

assumed that the user makes their own selection of

algorithms and parametrization. If this knowledge is

available to the user, this will be an efficient approach

for both consumer and provider. However, if the

problem is novel or there is a lack of expertise which

algorithm will provide best results, there will be two

alternatives: first, the user executes an own search for

the best algorithm by trying different algorithms and

parametrizations. However, this will lead to high

costs due to the amount of needed computing

resources which may not be tolerated by the

consumer.

As a second alternative, the provider could take

care of the process to identify suitable solution

algorithms and parametrizations. In order to make a

good price to the consumer as well as minimizing

operational costs, this process should be very efficient

by utilizing economies of scale. Therefore, an

optimization knowledge base is proposed in the

workflow. This knowledge base includes (meta-)data

about different optimization problems as well as

about the effectivity and efficiency of solution

algorithms and their parametrization. While the

knowledge base is sparse, a high number of

experiments should be done by the provider in order

to generate this data. In order to keep response times

low, efficient parallelization should be applied.

Although this will induce high operational costs in the

beginning, these can be dramatically reduced when

the knowledge base is enriched.

While this idea is novel to the field of meta-

heuristic optimization, such concepts have been

developed for the field of machine learning in which

the question which algorithm to use with which hyper

parameters is also a difficult one (meta-learning).

These include landmarking (Pfahringer et al. 2000) or

Towards an Automated Optimization-as-a-Service Concept

341

meta-regression (Charles et al. 2000; García-Saiz and

Zorilla 2017).

When a method has been selected and

parametrized, the experiment can be executed by

applying the optimization method on the decision

variables and the defined objective function to

generate results. If the user is not satisfied with these

results, they can select another one or with the gained

knowledge, the optimization method can be chosen

more carefully. Otherwise, results are returned to the

consumer.

4 ILLUSTRATION: THE

REDUNDANCY ALLOCATION

PROBLEM

In order to illustrate the generic workflow description

given above, the redundancy allocation problem

(RAP) is used. In this problem, the allocation of

parallel-redundant components in a series system is to

be optimized. The following, simple problem

definition is reported to be NP-hard (Chern 1992): A

system consists of  required subsystems. In each

subsystem, a number of components can be operated

in active redundancy. These components are

characterized by a reliability 



and cost 



. The

reliability of the system is to be maximized subject to

a cost constraint. With respect to the generic problem

definition given above, this would lead to:

 Decision variables 



   



with 







   





indicating the number of

components to be allocated in each subsystem

(limited by an upper bound),

 As a constraint, system cost should not exceed a

certain value:















 , and,

 The objective function under the assumption of

independent component failures, which is a non-

linear function





 



  















However, this formulation has been adapted and

extended in the last decades in order to approach

reality of complex systems. For instance,

heterogeneous or passive redundancy have been

introduced in decision variables (Coit and Smith

1996; Sadjadi and Soltani 2015). This led to more

complex objective functions that would be evaluated

by white-box, e.g. in (Bosse et al. 2016; Chi and Kuo

1990; Lins and Droguett 2009), or black-box

approaches, e.g. in (Hoffmann et al. 2004; Silic et al.

2014). Therefore, a possible user of an RAP module

in an OaaS context would require to freely define the

problem class to be solved.

Several meta-heuristic algorithms have been

developed in recent years to solve different RAP.

These include, for instance, genetic algorithms,

simulated annealing, tabu, harmony and cuckoo

search, as well as ant colony, immune-based, swarm,

and bee colony optimization (Soltani 2014).

Although several experiments have been conducted

and presented in the literature, the question which

algorithm is to be preferred under which

circumstances remains unanswered in a general scale

(Kuo and Prasad 2000). Additionally, algorithm

selection and parametrization can depend on the exact

problem formulation.

As an example, consider the boundary for the

number of components 



. This boundary has been

intended to limit the problem space in order to

increase efficiency of solution algorithms. However,

it has also an effect to parametrization as illustrated

by the genetic algorithm presented in (Coit and Smith

1996): In this paper, a solution is encoded as an

integer string of length   



, in which every

integer indicating the index of the component used or,

if the integer is the successor of the last index, that no

component is used in this slot. If 



is a large

number or is even undefined, this encoding scheme

cannot be applied effectively in a genetic algorithm,

so that other encoding schemes should be utilized.

In order to efficiently offer an OaaS, the provider

would require to run many experiments to serve

requests effectively. In these experiments, different

algorithms and parametrizations are to be applied to a

specific RAP. By relating information about problem

formulation (e.g. number of decision variables, upper

bound, ratio of cost constraint to mean component

cost etc.) and class (e.g. type of objective function) to

quality of solution algorithms (e.g. solution feasible,

(penalized) objective value etc.), the RAP knowledge

base can be filled and analyzed. On this basis, the

provider can select and parametrize solution methods

more efficiently for future requests.

5 CONCLUSION AND FUTURE

WORK

Combining meta-heuristics and cloud computing

would allow a high number of organizations to

leverage the opportunities of prescriptive analytics

without obtaining dedicated resources or expert

knowledge. While some concepts for an

optimization-as-a-service exist, these are not

discussing the workflow challenges of such a self-

service. In order to achieve a high degree of

IoTBDS 2019 - 4th International Conference on Internet of Things, Big Data and Security

342

automation, especially the smart selection and

parametrization of methods should be the focus of

future research. This would allow a provider to

minimize operational costs and guaranteeing low

services prices even for organizations without

optimization capabilities.

REFERENCES

Bosse, S., Splieth, M. and Turowski, K., 2016. Multi-

Objective Optimization of IT Service Availability and

Costs. Reliability Engineering & System Safety, 147,

pp.142–155. Available at:

http://www.sciencedirect.com/science/article/pii/S095

1832015003312.

Charles, C.K., Taylor, C. and Keller, J., 2000. Meta-

analysis: From data characterisation for meta-learning

to meta-regression. In Proceedings of the PKDD-00

workshop on data mining, decision support, meta-

learning and ILP.

Chern, M.-S., 1992. On the computational complexity of

reliability redundancy allocation in a series system.

Operations Research Letters, 11, pp.309–315.

Chi, D.-H. and Kuo, W., 1990. Optimal Design for

Software Reliability and Development Cost. IEEE

Journal on Selected Areas in Communications, 8(2),

pp.276–282.

Coit, D.W. and Smith, A.E., 1996. Reliability Optimization

of Series-Parallel Systems Using a Genetic Algorithm.

IEEE Transactions on Reliability, 45, pp.254–266.

Evans, J.R. and Lindner, C.H., 2012. Business analytics: the

next frontier for decision sciences. Decision Line,

43(2), pp.4–6.

Foster, I. et al., 2008. Cloud Computing and Grid

Computing 360-Degree Compared. 2008 Grid

Computing Environments Workshop, abs/0901.0(5),

pp.1–10.

García-Saiz, D. and Zorilla, M., 2017. A meta-learning

based framework for building algorithm

recommenders: An application for educational area.

Journal of Intelligent and Fuzzy Systems, 32, pp.1449–

1459.

Gill, P.E., Murray, W. and Wright, M.H., 1993. Practical

Optimization, Academic Press.

Hoffmann, G.A., Salfner, F. and Malek, M., 2004.

Advanced Failure Prediction in Complex Software

Systems, Informatik-Bericht 172 der Humboldt-

Universität zu Berlin.

King, J. and Magoulas, R., 2015. 2015 Data Science Salary

Survey, O’Reilly Media. Available at:

https://duu86o6n09pv.cloudfront.net/reports/2015-

data-science-salary-survey.pdf.

Kuo, W. and Prasad, V.R., 2000. An Annotated Overview

of System-Reliability Optimization. IEEE Transactions

on Reliability, 49(2), pp.176–187.

Kurschl, W. et al., 2014. Concepts and Requirements for a

Cloud-based Optimization Service. In Asia-Pacific

Conference on Computer Aided System Engineering

(APCASE).

Lanza, G., Haefner, B. and Kraemer, A., 2015.

Optimization of selective assembly and adaptive

manufacturing by means of cyber-physical system

based matching. CIRP Annals, 64(1), pp.399–402.

Lins, I.D. and Droguett, E.L., 2009. Multiobjective

optimization of availability and cost in repairable

systems design via genetic algorithms and discrete

event simulation. Pesquisa Operacional, 29, pp.43–66.

Marston, S. et al., 2011. Cloud Computing - The Business

Perspective. In 44th Hawaii International Conference

on System Sciences (HICSS).

Mell, P. and Grance, T., 2011. The NIST Definition of

Cloud Computing. National Institute of Standards and

Technology - Special Publication, 800-145, pp.1–3.

Müller, H., Bosse, S. and Turowski, K., 2016. Optimizing

server consolidation for enterprise application service

providers. In Proceedings of the 2016 Pacific Asia

Conference on Information Systems.

Nahhas, A. et al., 2017. Metaheuristic and hybrid

simulation-based optimization for solving scheduling

problems with major and minor setup times. In A. G.

Bruzzone et al., eds. 16th International Conference on

Modeling and Applied Simulation (MAS). Rende, Italy.

Pfahringer, B., Bensusan, H. and Giraud-Carrier, C.G.,

2000. Meta-Learning by Landmarking Various

Learning Algorithms. In 17th International Conference

on Machine Learning (ICML). Stanford, CA, USA, pp.

743–750.

Pimminger, S. et al., 2013. Optimization as a Service: On

the Use of Cloud Computing for Metaheuristic

Optimization. In R. Moreno-Diaz, F. Pichler, & A. Q.

Arencibia, eds. 14th International Conference on

Computer-Aided Systems Theory (EUROCAST).

Lecture Notes in Computer Science. Las Palmas De

Gran Canaria, Spain, pp. 348–355.

Pohl, M., Bosse, S. and Turowski, K., 2018. A Data-

Science-as-a-Service Model. In 8th International

Conference on Cloud Computing and Service Science

(CLOSER).

Sadjadi, S.J. and Soltani, R., 2015. Minimum–Maximum

regret redundancy allocation with the choice of

redundancy strategy and multiple choice of component

type under uncertainty. Computers & Industrial

Engineering.

Silic, M. et al., 2014. Scalable and Accurate Prediction of

Availability of Atomic Web Services. IEEE

Transactions on Service Computing, 7(2), pp.252–264.

Soltani, R., 2014. Reliability optimization of binary state

non-repairable systems: A state of the art survey.

International Journal of Industrial Engineering

Computations, 5, pp.339–364.

Towards an Automated Optimization-as-a-Service Concept

343