Machine Learning Algorithm in Period Estimation

1
Machine Learning Algorithm inPeriod Estimation

• Min-Su Shin
• Department of Astronomy, Yonsei University
• and
• Dongseon Kim Kiseok Do
• Department of Computer Science, Sogang University

2
Process of period estimation I (review)
Function fitting method

• Phase Dispersion Minimization method
• Orthogonal Complex Polynomial Function Fitting
  method

PDM method
3
Process of period estimation II (review)

• The sum of Cubic Spline functions is fitted to
  folded time-series data.

4
Problem is
What is the most appropriate period of the
variable star?
Discrimination of the most appropriate
period from spurious periods by a machine
learning algorithm
5
Machine learning algorithm

• Intelligent behavior requires knowledge.
• If we can program computers to learn from
  experience, we can bring various tasks within
  automation.

Learning Inference Memorization
6
Usage of a machine learning algorithm in other
cases in astronomy

• Classification of spectrum

• Determination of morphology

7
Light curve and machine learning (1)
8
Light curve and machine learning (2)
9
Experiment for determination of periods

• Selection of the software WEKA
• Input parameters from a light curve are
• Flat portion
• Dispersion in mag.
• Dispersion in phase
• of the local dim points
• of repeated local dim points
• of repeated local bright points
• Slope around local dim points
• Max interval in phase
• Used Algorithms are ZeroR, ID3, and etc.

10
Pre-processing

• Noise elimination by the gravity model
• Extraction of features from the noiseless light
  curve

11
Result of a simple experiment

• Among 300 samples, 1/3 is a true sample. Others
  are false samples.
• The Prism algorithm showed the highest
  performance!

12
Preliminary analysis

• The Prism algorithm is one of the rule-based
  algorithms, which constructs if-then rules for
  learning structure.
• When complex rule is adopted, the performance is
  generally improved.

13
Problems

• Noise elimination deforms the shape of a light
  curve.
• No detection of some
• features

14
Future plan

• What are appropriate features of light curves?
• Without preprocessing, image pattern recognition
  and classification?
• I need more help from computer scientists in
  using machine learning technologies.
• But I hope

15
Hunting for new variable stars in 2004
The discovery of YSTAR in December 2003?