SoftLabBoğaziçi University Department of Computer Engineering

Software Engineering Research Labhttp://softlab.boun.edu.tr/

Research Challenges Trend to large, heterogenous,

distributed sw systems leads to an increase in system complexity

Software and service productivity lags behind requirements

Increased complexity takes sw developers further from stakeholders

Importance of interoperability, standardisation and reuse of software increasing.

Research Challenges Service Engineering Complex Software Systems Open Source Software Software Engineering Research

Software Engineering Research Approaches

Balancing theory and praxis How engineering research differs

from scientific research The role of empirical studies Models for SE research

The need to link research with practice

Why after 25 years of SE has SE research failed to influence industrial practice and the quality of resulting software?

Potts argues that this failure is caused by treating research and its application by industry as separate, sequential activities.

What he calls the research-then-transfer approach. The solution he proposes is the industry-as-laboratory approach.

.

Colin Potts, Software Engineering Research Revisited, IEEE Software, September 1993

Research-then-Transfer

Incremental Refinementof research solutions

Problem evolves invisibly to the research community

ProblemV1

Wide gulf bridgedby indirect, anecdotal knowledge

ResearchSolutionV1

Technologytransfer Gapbridged byhard, but frequentlyinappropriate technology

ProblemV2

ProblemV3

ProblemV4

ResearchSolutionV2

ResearchSolutionV3Research

SolutionV4

Research-then-Transfer Problems Both research and practice evolve

separately Match between current problems in

industry and research solutions is haphazard

No winners

Disadvantages of Research-then-Transfer

Research problems described and understood in terms of solution technology - whatever is current research fashion. Connection to practice is tenuous.

Concentration is on technical refinement of research solution - OK but lacks industrial need as focus, so effort may be misplaced.

Evaluation is difficult as research solutions may use technology that is not commonly used in industry

Delay in evaluation means problem researchers are solving has often evolved through changes in business practice, technology etc.

Transfer is difficult because industry has little basis for confidence in proposed research solution.

Industry-as-Laboratory Approach to SE research

ProblemV2

Problem

V1

Problem

V3

Problem

V4

ResearchSolution

V1

ResearchSolution

V2

ResearchSolution

V3

ResearchSolution

V4

Advantages of Industry-as-Laboratory Approach

Stronger connection at start because knowledge of problem is acquired from the real practitioners in industry, often industrial partners in a research consortium.

Connection is strengthened by practitioners and researchers constantly interacting to develop the solution

Early evaluation and usage by industry lessens the Technology Transfer Gap.

Reliance on Empirical Research shift from solution-driven SE to problem-focused SE solve problems that really do matter to practitioners

Early SEI industrial survey research

What a SEI survey* learned from industry: There was a thin spread of domain knowledge in most

projects Customer requirements were extremely volatile.

These findings point towards research combining work on requirements engineering with reuse - instead of the approach of researching these topics by separate SE research communities - as is still found today!

*From ‘A field study of the Software Development Processfor Large Systems’, CACM, November 1988.

Further Results from Potts et al Early 90s Survey23 software development organizations (during 1990-92).

(Survey focused on Requirements Modeling process) Requirements were invented not elicited. Most development is maintenance. Most specification is incremental. Domain knowledge is important. There is a gulf between the developer and user User-interface requirements continually change. There is a preference for office-automation tools

over CASE tools to support development. I.e. developers found using a WP + DB more useful than any CASE tools.

Industry-as-Laboratory emphasizes Real Case StudiesAdvantages of case studies over studying

problems in research lab. Scale and complexity - small, simple (even simplistic)

cases avoided - these often bear little relation to real problems.

Unpredictability - assumptions thrown out as researchers learn more about real problems

Dynamism - a ‘real’ case study is more vital than a textbook account

The real-world complications of industrial case studies are more likely to throw up representative problems and phenomena than research laboratory examples influenced by the researchers’ preconceptions.

Need to consider Human/Social Context in SE research

Not all solutions in software engineering are solely technical.

There is a need to examine organizational, social and cognitive factors systematically as well.

Many problems are “people problems”, and require “people-orientated” solutions.

Theoretical SE research

While there is still a place for innovative, purely speculative research in Software Engineering, research which studies real problems in partnership with industry needs to be given a higher profile.

These various forms of research ideally complement one another.

Neither is particularly successful if it ignores the other. Too industrially focused research may lack adequate

theory!

Academically focused research may miss the practice!

Research models for SE Problem highlighted by Glass*: Most SE Research in 1990s was Advocacy Research.

Better research models needed. The software crisis provided the platform on

which most 90s research was founded. SE Research ignored practice, for the most part;

lack of practical application and evaluation were gapping holes in most SE research.

Appropriate research models for SE are needed.* Robert Glass, The Software -Research Crisis, IEEE Software, November 1994

Methods underlying Models

Scientific method

Engineering method

Empirical method

Analytical methodFrom W.R.Adrion, Research Methodology in

Software Engineering, ACM SE Notes, Jan. 1993

Scientific method

Observe real world

Propose a model or theoryof some real world phenomena

Measure and analyzeabove

Validate hypotheses of the model or theory

If possible, repeat

Engineering methodObserve existing solutions

Propose better solutions

Build or develop bettersolution

Measure, analyze, andevaluate

Repeat until no further improvements are possible

Empirical methodPropose a model

Develop statistical or otherbasis for the model

Apply to case studies

Measure and analyze

Validate and then repeat

Analytical methodPropose a formal theory or set of

axioms

Develop a theory

Derive results

If possible, compare withempirical observations

Refine theory if necessary

Need to move away from purely analytical method

The analytical method was the most widely used in mid-90s SE research, but the others need to be considered and may be more appropriate in some SE research.

Good research practice combines elements on all these approaches.

4 important phases for any SE research project (Glass)

Informational phase - Gather or aggregate information via reflection literature survey people/organization survey case studies

Propositional phase - Propose and build hypothesis, method or algorithm, model, theory or solution

Analytical phase - Analyze and explore proposal leading to demonstration and/or formulation of principle or

theory Evaluation phase - Evaluate proposal or analytic findings by

means of experimentation (controlled) or observation (uncontrolled, such as case study or protocol analysis) leading to a substantiated model, principle, or theory.

Software Engineering Research Approaches

The Industry-as-Laboratory approach links theory and praxis

Engineering research aims to improve existing processes and/or products

Empirical studies are needed to validate Software Engineering research

Models for SE research need to shift from the analytic to empirical.

Empirical SE Research

SE Research Intersection of AI and Software

Engineering An opportunity to:

Use some of the most interesting computational techniques to solve some of the most important and rewarding questions

AI Fields, Methods and Techniques

What Can We Learn From Each Other?

Software Development Reference Model

Intersection of AI and SE Research

Empirical Software Engineering

Intersection of AI and SE Research

Build Oracles to predict Defects Cost and effort Refactoring

Measure Static code attributes Complexity and call graph structure

Data collection Open repositories (NASA, Promise) Open source Softlab Data Repository (SDR)

Software Engineering Domain Classical ML applications

Data miner performance The more data the better the

performance Little or no meaning behind the

numbers, no interesting stories to tell

Software Engineering Domain Algorithm performance Understanding Data

Change training data: over/ under/ micro sampling Noise analysis Increase information content of data Feature analysis/ weighting Learn what you will predict later Cross company vs within company data

Domain Knowledge SE ML

In Practise Product quality

Lower defect rates Less costly testing times Low maintenance cost

Process quality Effort and cost estimation Process improvement

Software Engineering Research

Predictive Models Defect prediction and cost estimation Bioinformatics

Process Models Quality Standards Measurement

Major Research Areas Software Measurement

Defect Prediction/ Estimation

Effort & Cost Estimation

Process Improvement (CMM)

Defect Prediction Software

development lifecycle:

Requirements Design Development Test (Takes ~50% of overall

time) Detect and correct

defects before delivering software.

Test strategies: Expert judgment Manual code reviews Oracles/ Predictors as

secondary tools

Dynamicanalyser

Programbeing tested

Testresults

Testpredictions

Filecomparator

Executionreport

Simulator

Sourcecode

Testmanager Test data Oracle

Test datagenerator

Specification

Reportgenerator

Test resultsreport

A Testing Workbench

Static Code Attributes void main() { //This is a sample code

//Declare variables int a, b, c;

// Initialize variables a=2; b=5;

//Find the sum and display c if greater than zero

c=sum(a,b); if c < 0 printf(“%d\n”, a); return; }

int sum(int a, int b) { // Returns the sum of two

numbers return a+b; }

Module LOC LOCC V CC Error

main() 16 4 5 2 2

sum() 5 1 3 1 0

LOC: Line of CodeLOCC: Line of commented CodeV: Number of unique operands&operatorsCC: Cyclometric Complexity

c > 0

c

Defect Prediction

Machine Learning based models. Defect density estimation Defect prediction between versions Defect prediction for embedded systems

“Software Defect Identification Using Machine Learning Techniques”, E. Ceylan, O. Kutlubay, A. Bener, EUROMICRO SEAA, Dubrovnik, Croatia, August 28th - September 1st, 2006"Mining Software Data", B. Turhan and O. Kutlubay, Data Mining and Business Intelligence Workshop in ICDE'07 , İstanbul, April 2007 "A Two-Step Model for Defect Density Estimation", O. Kutlubay, B. Turhan and A. Bener, EUROMICRO SEAA, Lübeck, Germany, August 2007

“Defect Prediction for Embedded Software”, A.D. Oral and A. Bener, ISCIS 2007, Ankara, November 2007

"A Defect Prediction Method for Software Versioning", Y. Kastro and A. Bener, Software Quality Journal (in print). “Ensemble of Defect Predictors: An Industrial Application in Embedded Systems Domain.” Tosun, A., Turhan, B., Bener, A. A, and Ulgur, N.I., ESEM 2008.B.Turhan, A. Tosun and A. Bener, "An Industrial Application of Classifier Ensembles for Locating Software Defects". Submitted to Information and Software Technology Journal, 2008.

Constructing Predictors Baseline: Naive Bayes. Why?: Best reported results so far (Menzies et al., 2007) Remove assumptions and construct different models.

Independent Attributes ->Multivariate dist. Attributes of equal importance -> Weighted Naive Bayes

"Software Defect Prediction: Heuristics for Weighted Naïve Bayes", B. Turhan and A. Bener, ICSOFT2007, Barcelona, Spain, July 2007.“Software Defect Prediction Modeling”, B. Turhan, IDOESE 2007, Madrid, Spain, September 2007“Yazılım Hata Kestirimi için Kaynak Kod Ölçütlerine Dayalı Bayes Sınıflandırması”, UYMS2007, Ankara, September 2007“A Multivariate Analysis of Static Code Attributes for Defect Prediction”, B. Turhan and A. Bener QSIC 2007, Portland, USA, October 2007.“Weighted Static Code Attributes for Defect Prediction”, B.Turhan and A. Bener, SEKE 2008, San Francisco, July 2008.B.Turhan and A. Bener, "Analysis of Naive Bayes' Assumptions on Software Fault Data: An Empirical Study". Data and Knowledge Engineering Journal, 2008, in printB.Turhan, A. Tosun and A. Bener, "An Industrial Application of Classifier Ensembles for Locating Software Defects". Submitted to Data and Knowledge Engineering Journal, 2008.B.Turhan, A. Bener and G. Kocak "Data Mining Source Code for Locating Software Bugs: A Case Study in Telecommunication Industry". Submitted to Expert Systems with Applications Journal, 2008.

WC vs CC Data for Defects? When to use WC or CC? How much data do we need to

construct a model?

“Implications of Ceiling Effects in Defect Predictors”, Menzies, T., Turhan, B., Bener, A., Gay, G., Cukic, B., Jiang, Y. PROMISE 2008, Leipzig, Germany, May 2008.“Nearest Neighbor Sampling or Cross Company Defect Predictors”, Turhan, B., Bener, A., Menzies, T., DEFECTS 2008, Seattle, USA, July 2008."On the Relative Value of Cross-company and Within-Company Data for Defect Prediction", B. Turhan, T. Menzies, A. Bener, J. Distefano, Empirical Software Engineering Journal, 2008, in printT. Menzies, Z.Milton, B. Turhan, Y. Jiang, G. Gay, B. Cukic, A. Bener, "Overcoming Ceiling Effects in Defect Prediction", Submitted to IEEE Transactions on Software Engineering, 2008.

Module Structure vs Defect Rate

Fan-in, fan-out Page Rank Algorithm Dependency graph information “small is beautiful”

Koçak, G., Turhan, B., Bener,A. “Software Defect Prediction Using Call Graph Based Ranking Algorithm”, Euromicro 2008.

G. Kocak, B. Turhan and A.Bener, "Predicting Defects in a Large Telecommunication System”, ICSOFT'08.

COST ESTIMATION Cost Estimation: predicting the effort effort required to

develop a new software project Effort: the number of months one person would

need to develop a given project (person months-PM) CE assists project managers when they make

important decisions (bidding, planning, resource allocation):

underestimation approve projects that would then exceed their budgets

overestimation waste of resources Modeling accurate & robust cost estimators =

Successful software project management

COST ESTIMATION Understanding the data structure?

CROSS- vs. WITHIN-APPLICATION DOMAIN embedded software domain

Better predictor? Point Estimation: a single value of effort is

tried to be estimated Interval Estimation: effort intervals are

tried to be estimatedCOST CLASSIFICATION dynamic intervals

classification algorithmspoint estimates

COST ESTIMATION

How can we achieve accurate estimations with limited amount of effort data?

feature subset selection: Save the cost of extracting less important features

KNNAlgorithm

SimilarProject

1

ENN

EstimatedEffort

(1)

EstimatedEffort

(k)

EstimatedEffort

(2) BiasCalculation

ActualEffort

(1)

ActualEffort

(k)

ActualEffort

(2)

PastProjects

EstimatedBias

+Estimated

EffortENNA

SimilarProject

k

SimilarProject

2

NewProject

Cost Estimation Comparison of ML based models with parametric models Feature ranking COCOMO81- COCOMO2-COQUALMO Cost estimation as a classification problem (interval

prediction)

"Mining Software Data", B. Turhan and O. Kutlubay, Data Mining and Business Intelligence Workshop in ICDE'07 , İstanbul, April 2007“Software Effort Estimation Using Machine Learning Methods”, B. Baskeles, B.Turhan, A. Bener, ISCIS 2007,Ankara, November 2007."Evaluation of Feature Extraction Methods on Software Cost Estimation", B. Turhan, O. Kutlubay, A. Bener, ESEM2007, Madrid, Spain, September 2007 . “ENNA: Software Effort Estimation Using Ensemble of Neural Networks with Associative Memory” Kültür Y., Turhan B., Bener A., FSE 2008.“Software Cost Estimation as a Classification Problem”, Bakır, A., Turhan, B., Bener, A. ICSOFT 2008.B.Turhan, A. Bakir and A. Bener, "A Comparative Study for Estimating Software Development Effort Intervals". Submitted to Knowledge Based Systems Journal, 2008. B.Turhan, Y. Kultur and A. Bener, "Ensemble of Neural Networks with Associative Memory (ENNA) for Estimating Software Development Costs", Submitted to Knowledge Based Systems Journal, 2008. A. Tosun, B. Turhan, A. Bener, "Feature Weighting Heuristics for Analogy Based Effort Estimation Models", Submitted to Expert Systems with Applications, 2007. A. Bakir, B.Turhan and A. Bener, "A New Perspective on Data Homogeneity for Software Cost Estimation". Submitted to Software Quality Journal, 2008.

Prest

A tool developed by Softlab

Parser C, Java, C++, jsp

Metric Collection Data Analysis

Public Datasets NASA (IV&V Facility, Metrics Program) PROMISE (Software Engineering Repository)

Includes Softlab data now Open Source Projects (Sourceforge, Linux, etc.) Internet based small datasets University of South California (USC) Dataset Desharnais Dataset ICBSG Dataset NASA COCOMO and NASA 93 Datasets

Softlab Data Repository (SDR) Local industry collaboration Total 20 companies, 25 projects over 5 years

Data Sources

Process Automation UML Refactoring

Class diagram – source code Tool Algorithm (graph based)

What needs to be refactored Complexity vs call graphs

Y. Kösker and A. Bener . "Synchronization of UML Based Refactoring with Graph Transformation", SEKE 2007, Boston, July 9-11, 2007 B.Turhan, Y. Kosker and A. Bener, "An Expert System for Determining Candidate Software Classes for Refactoring". Submitted to Expert Systems with Applications Journal, 2008.Y. Kosker, A.Bener and B. Turhan, "Refactoring Prediction Using Class Complexity Metrics”, ICSOFT'08, 2008.B. Turhan, A. Bener and Y.Kosker, "Tekrar Tasarim Gerektiren Siniflarin Karmasiklik Olcutleri Kullanilarak Modellenmesi" (in Turkish), 2. Ulusal Yazilim Mimarisi Konferansi (UYMK'08), 2008.

Process Improvement and Assessment

A Case in health care industry Process Improvement with CMMI

Requirements Management Change Management

Comparison: A Before and After Evaluation

Lessons LearnedA. Tosun, B. Turhan and A. Bener,"The Benefits of a Software Quality Improvement Project in a Medical Software Company: A Before and After Comparison", Invited Paper and Keynote speech in International Symposium on Health Informatics and Bioinformatics (HIBIT'08), 2008.

Metrics Program in Telecom Metrics extraction from 25 Java applications

Static code attributes (McCabe, Halstead and LOC metrics) CallGraph information (caller-callee relation between modules) Information from four versions (a version in two weeks) Product and test defects (pre-release defects)

Various experimental designs for predicting fault-prone files

Discard versions: Treat all applications in a version as a single project Predict fault-prone parts of each application

Using previous versions of all the applications Using previous versions of the selected application

Additionally, Optimization of the local prediction model using CallGraph metric Refactoring prediction using class complexity metrics

Requirements Analysis

Coding

Design

Test

Maintenance ∞

Defect prediction

Call Graph / Refactoring

Refactoring

Matching reqs with defects

Test driven development

Emerging Research Topics Adding organizational factors to local prediction model

Information about the development team, experience, coding practices, etc. Adding file metrics from version history

Modified/added/deleted lines of code Selecting only modified files from each version in the prediction model

Confidence Factor Using time factors Dynamic prediction: Constructing a model

for each application in a version for each module/package in an application for each developer by learning from his/her coding habits

TDD Measuring test coverage Defect proneness Company wide implementation process Embedded systems

Cost/ Effort Estimation Dynamic estimation per process

Bioinformatics Investigating time-dependent behavior of embryo growth by using sequential data

analysis methods such as Hidden Markov Models SOA