Table of Contents

FuNN Tutorial

A. Creating the Network

  1. Use the following parameters to create your first FuNN neural network using FuzzyCope/3:

    Network Architecture:
    Number of Inputs 4
    Number of Members 3 (3 Memberhip Functions per input, eg High, Medium, Low)
    Number of Rules 5
    Number of Outputs 3
    Number of Actions 2 (2 Memberhip Functions per input, eg High, Low)

  2. Press the create button.

  3. Save the network as IrisFUNN.wgt

  4. View the network


B. Training the network

In neural net applications the weights are determined so that they minimise the Root Mean Squared errors (RMSE).

  1. Open the training file that you formatted earlier - Iris.trn

  2. View the training file

  3. Set the parameters for your Iris neural network:
    Learning Rate 0.3
    Momentum 0.6
    Epochs 100 then 500 (increase until the terminating error is reached)
    Terminating Error 0.01
    Training Mode Batch

  4. Press the Train button to commence training

  5. Save the trained network as TIrisFUNN.wgt


C. GA Training

  1. If your training error isn't going down try some GA (Genetic Algorithm) training. Use the following parameters:
    Population Size 50
    Minimum Weight -20
    Maximum Weight 20
    Mutation Rate 0.01
    Generations 100
    Terminating Error 0.01
    Fitness Normalisation Unchecked
    Elitism Checked
    Crossover points 1
    Selection strategy Tournament

    The training error will result.

  2. Save the trained network again as TIrisFUNN.wgt


D. Network Evaluation

  1. Choose the Recall Tab

  2. Open the recall file that you formatted earlier - Iris.rcl

  3. Open a new output file for the results and name it IrisFUNN.out

  4. Press the Recall Button.

  5. Save the new output file.

  6. View the Output file.

  7. Compare your results with the desired outputs


E. Extracting Rules

  1. Choose the Extract Rules Tab

  2. Press the label button

  3. Save the label file (using the Save As Label file option from the Extract Rules menu)

  4. Open the label file in a text editor and modify the labels

  5. Return to FuNN and load the modified label file

  6. Set the Threshold values
    Threshold In 1.0
    Threshold Out 1.0

  7. Press the Extract Rules button

  8. View your rule file (using the View Rule file option from the Extract Rules menu)


F. Additional Work

Try different combinations of:

  • Rules (Hidden Nodes)
  • Learning Rate
  • Momentum
  • Pattern / Batch Training Mode
  • Threshold values

    This page is maintained by Melanie Middlemiss mmiddlemiss@infoscience.otago.ac.nz
    Last modified on: 22/12/98.