Home Biografija Poslovni programi Rečnik PlanB Veštačka inteligen. Neuronske mreže ANN Force learn algorit. NeuroVCL Agenti ENAA OCR OCR soft Izrada SCADA SmartSCADA OPC standard Nanotehnologija Virtuelna učionica EDI Linkovi Free Download

Velibor Ilić - Web Page FORCE LEARN ALGRITHM (Training Paterns With Highest Error) Author: Velibor Ilić ABSTRACT: This study describes training methods of the neural nets with the backpropagation algorithm of learning in which the priority is given to the patterns of the training set with the maximum error during the training. Date: Mart, 2000 This work was persented on 5th seminar on neural networks (NEUREL) http://galeb.etf.bg.ac.rs/~NEUREL   Table of contents Introduction Force Learn algorithm Comparation Force Learn algorithm with regular method for training on particular example ANN V2.3 Neural Net training program References

Verzija na srpskom   Related with: OCR ANN NeuroVCL Artificial intelligence Neural networks Software agents ENAA

INTRODUCTION

The most common problem that appears during the neural net training is the “uneven training”. The neural net is quickly trained to recognized only some of the training set patterns, while the rest demand larger number of iterations for certain recognition. If some training set elements achieve almost 100% error, it’s difficult to reduce their errors in further training, even after relatively great number of iterations. The ideal solution would be if the errors were eventually reduced up to the allowed values during training.
 

FORCE LEARN ALGORITHM

On figure 1 is shown characteristic graph that is given by neural network training. On graph is followed errors (average and maximum) and percent correct outcome. By rectangle is marked critical period during neural network training. This is critical period because there is need for lot of number of iterations to provide maximum error for falling down. In this period maximum error has high value nearly or even equal 1, and percent of correct outcome is about 95-99% and further grows weakly or it doesn't grow.

Using force learn algorithm this situations can be avoided. Going out from critical period is faster because maximum error falls down much faster then using regular method of training.
The main reason for developing force learn algorithm was to increase probability that neural net could successfully learn the patterns of the training set, but the usage of this algorithm has achieved one very important effect, the reduction of time and iteration number necessary for training the net.

Algorithm consists of:
In one iteration instead of changing weights of connections for all elements of training set:

• first must be determinated value of variable training_limit as a maximum error from the previous iteration decreased for 20%,
• the elements of training set must be taken on input of neural net and error on output for each of this elements must be determinated,
• if error for particular element, that was given on output of neural net, is bigger then value of variable training_limit net has to be trained for particular example several time. In experience it has shown that 3 repeated training gives good results.

COMPARATION FORCE LEARN ALGORITHM WITH REGULAR METHOD FOR TRAINING ON PARTICULAR EXAMPLE

In this example neural network is trained to recognize 12 different patterns size 3×3 (figure 3.) in array size 5×5. These patterns can be placed in 9 positions in array (figure 2.). The training set for this net is consisted of 108 elements (the number of positions × the number of shapes) and the exact examples (figure 2 and 3). If there is a shape for which the net is not trained, later, in using the net after a successful training, the results in the interval from (limit error)<x<(1-limit error) will appear at the output.

This example is characteristic because it shows privilege of force learn algorithm. On figure 4 is shown graph of regular method for training neural net. Maximum error has reached value 1 and it keeps on that value. Percent of correct outcome has reached of 99% and it doesn't grow further. On figure 5, there is graph neural net with same configuration but trained with force learn algorithm and with random order of patterns on net input. Maximum error falls much faster despite lot of oscillations.

Neural network configuration:
Three layers neural net with backpropagation algorithm of training:
Number neurons at input layer (number of inputs): (figure 2.) 5×5 = 25
Number neurons at hidden layer: 20
Number neurons at output layer (number of outputs): 12
Learning coefficient (alpha): 0.25
Number of patterns in training set: 12×9 = 108 (Figure 2 and figure 3)

Training set:

a) Inputs:

figure 2. positions of object in array

figure 3. different objects

Network has 25 inputs and each of them can be added to one part of array. On figure 2., there are 9 positions, which can be taken in array by objects from figure 3.

b) output data:
One output of network is added to each of objects from figure that detects him. It means that on output of neural net appears one 1 and rest values are 0.
Type of processor: Intel Celeron 400
Type of program for training neural net: ANN V2.3

Figure 4. Normal order training – maximum and average error and percent of correct outcome

Number of iterations: 10000
Training time: 00:17:47
Average error: 0.01724
Maximum allowed error: 0.1
Maximum error: 1

Net was stopped in training on 10000-th iteration but on graph, on figure 4, there are only 2000 iterations, because the graph is almost same further. In that moment network was trained 99% (1281/1296 correct outputs). On figure 5., there is graph of net that was trained by using force learn algorithm for more less iterations.

Figure 5. Training using Force learn algorithm and random order – maximum and average error and percent of correct outcome

These graphs prove that maximum error reduces faster by application of force learn algorithm and random training methods, although there is more oscillation during the training.

Relations shown at the graphs 4 and 5 are only valid for this concrete example. If neural nets would be trained by another example, relation between normal order training and force learn method would be different from those shown at the graphs.

ANN V2.3 – NEURAL NET TRAINING PROGRAM

Figures 5 show the neural net training program display. Using this program, obtained results is shown at the graphs by previous graphs.

Figure 5. Neural net training program window

REFERENCES