Training RBFNs Using Differential Evolution

1
Training RBFNs Using Differential Evolution

• Junhong Liu
• Department of Information Technology
• Lappeenranta University of Technology
• P.O.Box 20, FIN-53851 Lappeenranta
• Email liu_at_lut.fi

2
Agenda

• Differential Evolution (DE)
• Performance
• Efficiency
• Function approximation using Radial Basis
  Function Networks (RBFNs)
• Further work

3
DE 1

• A floating-point encoded Evolutionary Algorithm
  for global optimization
• Problem Minimize f(X) ?D?? by finding a
  suitable parameter vector
• x (x1, x2, , xD)
• Control parameters F, Cr, and Npop

4
DE 2 Setting the Control Parameters

• A successful optimization process
• - How control parameter settings influence the
  process of DE (c1)
• Ways of improving the efficiency of DE (c 2-5,
  j1)

5
Function Approximation 1 Construct RBFNs

• Performance of a trained RBFN
• A variety of approaches
• Centres, widths
• Weights
• Orientations
• Learning the network structure

6
Function Approximation 2 A Typical RBFN
Configuration
7
Function Approximation 3

• Train RBFNs (c 7-9)
• Euclidean distance, MSE
• Variables centres, widths, and weights
• Fixed or variable number of RBFs
• Experiments test functions real data

8
Function Approximation 4

• A DE-based incremental training method
• Parameters centres, widths, and weights
• Studying control parameters setting
• Real-life data

9
Further Work

• No more courses
• Conduct the planed work
• Writing thesis since May