+ All Categories
Home > Documents > GECCO 2013 Industrial Competition

GECCO 2013 Industrial Competition

Date post: 23-Feb-2016
Category:
Upload: rosie
View: 30 times
Download: 0 times
Share this document with a friend
Description:
GECCO 2013 Industrial Competition. Computer Engineering Lab, School of Electrical and IT Engineering. Rommel Vergara. Introduction. Machine Learning Algorithm used: Kernel Recursive Least Squares - PowerPoint PPT Presentation
Popular Tags:
12
GECCO 2013 Industrial Competition Computer Engineering Lab, School of Electrical and IT Engineering Rommel Vergara
Transcript
Page 1: GECCO 2013 Industrial Competition

GECCO 2013 Industrial Competition

Computer Engineering Lab, School of Electrical and IT EngineeringRommel Vergara

Page 2: GECCO 2013 Industrial Competition

2

Introduction

›Machine Learning Algorithm used: Kernel Recursive Least Squares- Used an open-source C++ library, dlib (http://dlib.net) which has a full implementation of the KRLS algorithm.

›Challenges: - Aperiodic and missing data samples- Selection of the feature set to describe the Temperature and Humidity values.

Page 3: GECCO 2013 Industrial Competition

3

Data Preprocessing

›All data sets contained missing data and were aperiodic in nature.

›SOLUTION:- Missing Data: Missing data occurrences were linearly approximated between their adjacent available data points.

- Aperiodic Nature: Data was standardized into 10 minutes intervals (as required by the format for the competition)

Page 4: GECCO 2013 Industrial Competition

4

Data Preprocessing

›Data set was also narrowed to contain the target weekdays only: Tuesday, Wednesday and Thursday.

›Able to capture more accurately minor and specific changes that occur on these particular weekdays.

Page 5: GECCO 2013 Industrial Competition

5

Data Preprocessing

1 29 57 85 11314116919722525328130933736539342144947750553356158961764567370172975778581384120

20.5

21

21.5

22

22.5

23

23.5

24

24.5

Temperature (Tuesday, Wednesday, Thursday)

Temperature

Time (10 minute intervals)

Tem

pera

ture

Wed

ThuTue

Tue

WedThu

Page 6: GECCO 2013 Industrial Competition

6

Feature Set Selection

›The ‘Feature Set’ is a selection of inputs that contribute and explain the output

›Important to choose the feature set to describe the outputs we are predicting

›There are many ways to represent the feature set to describe an output.

›Each feature set is represented as a column vector and is fed into the KRLS algorithm.

Page 7: GECCO 2013 Industrial Competition

7

Temperature Feature Set

› The temperature feature set chosen contained 145 values- Current weather value- 144 temperature values 10 minute rolling window of the previous

and current weekday lagged by 1 week.

› REASONING: - Remove noise that exists in other weekdays- Allow KRLS to concentrate and focus more on the specific

weekdays that needed to be judged in the competition.

Page 8: GECCO 2013 Industrial Competition

8

Temperature Feature Set

› Example:- To predict the following data point:

- 20/02/2013 00:00 (Wednesday)

- The following temperature values were used:

- 12/02/2013 00:00 (Tuesday) to

- 12/02/2013 23:50 (Tuesday)

- To predict the next data point:

- 20/02/2013 00:10 (Wednesday)

- The following temperature values were used (10 minute rolling window):

- 12/02/2013 00:10 (Tuesday) to

- 13/02/2013 00:00 (Wednesday)

Page 9: GECCO 2013 Industrial Competition

9

Temperature Results

3 111 219 327 435 543 651 759 867 975 10831191129914071515162317311839194720552163227123792487259527032811291915

17

19

21

23

25

27

Temperature Time Series

PredictedActual

Time (10 minute intervals)

Tem

pera

ture

Page 10: GECCO 2013 Industrial Competition

10

Humidity Feature Set

› The humidity feature set chosen contained 2 values- Current weather value

- Predicted KRLS temperature value

› This proved to be ineffective, providing an RMSE of 0.12 in the competition.

› CHALLENGE: Humidity data set was aperiodic and observed to have discrete-like behaviour.

› IMPROVEMENT: Given a more continuous data set, I would have chosen the same technique as the Temperature Feature Set, which is to take a rolling window of previous and current weekday humidity values lagged by 1 week.

Page 11: GECCO 2013 Industrial Competition

11

Humidity Results

3 111 219 327 435 543 651 759 867 975 10831191129914071515162317311839194720552163227123792487259527032811291915

20

25

30

35

40

45

50

55

60

Humidity Time Series

PredictedActual

Time (10 minute intervals)

Hum

idity

Page 12: GECCO 2013 Industrial Competition

12

References

›“The Kernel Recursive Least Squares Algorithm” (2003), Yaakov Engel, Shie Mannor, Ron Meir.

›dlib library: http://dlib.net

›Contact:- Rommel Vergara (University Of Sydney)- [email protected]


Recommended