School of Electrical and Electronic Engineering

1
School of Electrical and Electronic
Engineering Information Communication Institute
of Singapore
Time Series Analysis by Radial Basis Function
Neural Networks
Objectives
Approach

• To improve the GAPRBF algorithm for
• Eliminating the undetermined conditions and
  parameters
• Having acceptable generalization performance
• Being applied in time series problems
  conveniently

• Analyze and study the related algorithms, include
  RAN, RANEKF, MRAN and GAPRBF methods
• Employing a probability rate to estimate the
  neurons significance in uniform distribution
  situation
• Depending on the approximation equation of
  significance, the overlap factor and distance
  condition could be removed away
• Using the distance between two hidden neurons
  center to calculate the input range
• DYNAMIC GAP-RBF ALGORITHM
• For each observation (xn , yn), do
• compute the overall network output and the error
• apply the criterion for adding neurons

Introduction

• Algorithms based on RBF structure are employed
  into many applications and obtain a satisfactory
  performance
• However, parameters in the algorithms
  predetermined by experiments limit the
  application in real industry environment. The
  parameters include overlap factor ,input range
  size S(x) and distance conditions
• On the other side, time series situation requires
  more convenient and effective algorithm
• Therefore, it is important to find a new
  algorithm which can remove the uncertain
  parameters and conditions away. In addition, the
  new algorithm can be used in time series
  environment

If
add a new neuron K1 with
Overlap Factor
else
adjust the network parameters for the nearest
neuron only, by EKF
endif
Distance Condition
Expected Results

• An efficient dynamic growing and pruning
  algorithm (DGAP-RBF) which can select parameters
  directly
• Applying the proposed DGAP-RBF algorithm to solve
  time series problems

Input Range Size
Student See Wang Ying
Supervisor Associate Professor Dr Huang Guangbin
Information Communication Institute of Singapore
? Blk S2, Level B3c-13, Nanyang Avenue,
Singapore 639798 ? Website http//www.icis.ntu.
edu.sg