16
School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus A-Class: a novel classification method I.Tsoulos, A. Tzallas, E. Glavas

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus A-Class: a novel classification method I.Tsoulos, A. Tzallas, E. Glavas

Embed Size (px)

Citation preview

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

A-Class: a novel classification method

I.Tsoulos, A. Tzallas, E. Glavas

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Presentation Layout

Data classification Grammatical evolutionMobile programmingImplementationExperimental resultsFuture work

2

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Data Classification Used in chemistry, economics, physics, medicine etc. Usually the data are divided into: Train data: A dataset used for the training of the proposed

method Test data: The dataset where the proposed method will be

evaluated Example of methods are: K-nearest neighbours Radial basis functions Artificial neural networks Support vector machines

3

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Grammatical Evolution

Genetic algorithm Introduced by Ryan and O'Neil It has been used in many scientific & practical

applications It requires:

The grammar of the target problem in BNF notation An associated fitness function

Our case… • the fitness is the classification error from the application of

the produced rules upon to the training set• Genetic Evolution is only used to transform a typical

chromosome into human readable programme

M. O’Neill, C. Ryan, Grammatical evolution, IEEE Trans. Evol. Comput. 5 (2001) 349–358

4

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Mobile programming

Our study is designed... not only in desktop environments

... can be executed in recent mobile devices Many programming languages:

Java for Android Objective C for Iphone C# for Windows Phones Javascript for Firefox OS

5

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Implementation

Use of QtCreatorUtilization of C++ languageFreely available from http://qt-project.org It can be installed in any operation system It can produce mobile applications for Android & IOS

6

This means… • We write our program once & the produced output can be

run in any mobile device• We can produce executables with the same source code in

any desktop environment

QtCreator Environment

7

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Algorithm Description

1. Read the train data of the problem2. Random initialization of the chromosomes3. For a number of generations Do• Fitness evaluation• Create a new genetic population using mutation &

crossover4. End-For5. Create a classification program induced by the best

chromosome in the population6. Apply the above program to the test set

8

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Experimental setup

Two (2) Datasets from UCI Repository Wine Glass

One (1) artificial dataset (Circular) Two fold Experiments (50 % train and 50% testing)30 individual runs for every dataset & averages are

taken

9

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

lif(x9>exp((947.6-(x13*log(x12))))) CLASS=0.00

else if(x10<=((-306.79)/(-87.77))|x7>=exp(x8)) CLASS=1.00

else

CLASS=2.00

Typical output of the software

Output for a random generation for the dataset wine

10

Screenshot of the execution of the method

11

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Results (1/2)DATASET GENERATIONS TEST ERROR

WINE 100 22.00%

WINE 500 17.11%

WINE 1000 15.30%

GLASS 100 43.46%

GLASS 500 41.40%

GLASS 1000 39.81%

CIRCULAR 100 25.08%

CIRCULAR 500 22.68%

CIRCULAR 1000 21.66%

Experiments using different number of generations & fixed size of chromosomes to 200

12

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Results (2/2)DATASET CHROMOSOMES TEST ERROR

WINE 200 17.11%

WINE 500 15.41%

WINE 1000 13.30%

GLASS 200 41.40%

GLASS 500 38.75%

GLASS 1000 37.32%

CIRCULAR 200 22.68%

CIRCULAR 500 19.89%

CIRCULAR 1000 17.50%

More experiments were conducted using fixed number of generations (set 500) & different number of chromosomes

13

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Conclusions

A novel method for classification problems…utilizes the Grammatical Evolution procedure to

create classification programs expressed in a C – like programming language

….was tested on a series of well known problemsThe associated software was implemented using Qt

Creator programming environment & was installed on Android mobile devices

14

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Future Work

The software can be extended in the following ways: Implementation & inclusion of a better stopping rule• Currently, the software terminates using a maximum number of

generations • This is not efficient & it can consume the battery of the mobile

device very fast in some cases Addition of a new button to access program settings Support a better mechanism of fetching datasets Application to real world problems from areas such as

medicine & economics

15

School of Applied Technology, Dep. Of Computer Engineering, T.E.I of Epirus

Thank you!!!

16