Controls

Controls
 
 
When a new neural network is created the controls are preset to values that will usually be suitable for the network to learn from the training data in the Grid. The training controls can often be optimized to result in much faster learning.  All controls are set using the Controls dialog. 
 
Setting the Learning Rate
 
The Learning Rate will be preset to 0.6 when a new neural network is created.  It can be changed to any value from 0.1 to 10.  Very low values will result in slow learning and values above 1.5 will often result in erratic learning or oscillations.  Check "Optimize" to allow EasyNN-plus to determine the learning rate automatically by running a few learning cycles with different learning rate values.  Check "Decay" to automatically reduce the learning rate during learning if erratic learning or oscillations occurs.  Check "Adapt" to maintain the learning rate that produces the fastest learning.
 
Setting the Momentum
 
The Momentum will be preset to 0.8 when a new neural network is created.  It can be changed to any value from 0 to 0.9.  Check "Optimize" to allow EasyNN-plus to determine the momentum automatically by running a few learning cycles with different momentum values.  Check "Decay" to automatically reduce the momentum during learning if oscillations occur. Check "Adapt" to maintain the momentum that produces the fastest learning.
 
Setting the Accelerator
 
The Accelerator will be preset to 0.0 when a new neural network is created.  It can be changed to any value from 0 to 3.  Check "Optimize" to allow EasyNN-plus to determine the accelerator automatically. Check "Decay" to automatically reduce the accelerator during learning if oscillations occur.
 
Setting the Target Error
 
The Target Error will be preset to 0.01 when a new neural network is created.  It can be changed to any value from 0 to 0.9 but values above 0.2 usually result in under trained networks.  Learning is set to stop when the average error drops below the target error.  This can be changed so that all errors must drop below the target or the predicted output is in range of the actual output.
 
Setting the Validating Controls
 
Validating cycles only occur when validating examples are included in the Grid.  A number of training examples can be randomly changed to validating examples using the Select control or they can be entered directly into the Grid.  The cycle values are preset so that 100 learning cycles occur for every validating cycle.  The number of learning cycles before the first validating cycles and the number of learning cycles per validating cycle can be changed to any positive values.  Learning can be set to stop when the validating target is reached.  The validating target can be set to any value up to 100%.  The validating test can be either when the validating results are within a specific range or when the results are correct after rounding to the nearest whole number.  Validating Rules can be used for rounding to the nearest multiple.  The validating range can be set to any value from 0 to 50%.
1

Learning

Set the learning rate, momentum and accelerator manually or by finding the optimum values.  The learning rate, momentum and accelerator can be set to decay during learning if the descent to the target error is irregular.  The learning rate and momentum can be set to adapt to maintain the fastest learning.  The number of learning threads can be set to take advantage of multiprocessor systems.
 
 
 
When a new neural network is created the controls are preset to values that will usually be suitable for the network to learn from the training data in the Grid. The training controls can often be optimized to result in much faster learning.  All controls are set using the Controls dialog. 
 
Setting the Learning Rate
 
The Learning Rate will be preset to 0.6 when a new neural network is created.  It can be changed to any value from 0.1 to 10.  Very low values will result in slow learning and values above 1.5 will often result in erratic learning or oscillations.  Check "Optimize" to allow EasyNN-plus to determine the learning rate automatically by running a few learning cycles with different learning rate values.  Check "Decay" to automatically reduce the learning rate during learning if erratic learning or oscillations occurs.
 
Setting the Momentum
 
The Momentum will be preset to 0.8 when a new neural network is created.  It can be changed to any value from 0 to 0.9.  Check "Optimize" to allow EasyNN-plus to determine the momentum automatically by running a few learning cycles with different momentum values.  Check "Decay" to automatically reduce the momentum during learning if oscillations occur.
 
Setting the Accelerator
 
The Accelerator will be preset to 0.0 when a new neural network is created.  It can be changed to any value from 0 to 3.  Check "Optimize" to allow EasyNN-plus to determine the accelerator automatically. Check "Decay" to automatically reduce the accelerator during learning if oscillations occur.
 
Setting the Target Error
 
The Target Error will be preset to 0.01 when a new neural network is created.  It can be changed to any value from 0 to 0.9 but values above 0.2 usually result in under trained networks.  Learning is set to stop when the average error drops below the target error.  This can be changed so that all errors must drop below the target or the predicted output is in range of the actual output.
 
Setting the Validating Controls
 
Validating cycles only occur when validating examples are included in the Grid.  A number of training examples can be randomly changed to validating examples using the Select control or they can be entered directly into the Grid.  The cycle values are preset so that 100 learning cycles occur for every validating cycle.  The number of learning cycles before the first validating cycles and the number of learning cycles per validating cycle can be changed to any positive values.  Learning can be set to stop when the validating target is reached.  The validating target can be set to any value up to 100%.  The validating test can be either when the validating results are within a specific range or when the results are correct after rounding to the nearest whole number.  Validating Rules can be used for rounding to the nearest multiple.  The validating range can be set to any value from 0 to 50%.
2

Network reconfiguration

2. Network reconfiguration  
Set manual or automatic hidden layer reconfiguration.
 
3

Validating

3. ValidatingSet validating cycles and selection of rows. Up to half of the training rows can be selected at random for validating.
 
4

Slow learning

4. Slow learningSet learning cycle delay for slow learning demonstration.
 
5

Stops

5. StopsStop when Average error is below the target error uses the total example errors divided by the number of examples.
 
Stop when All errors are below the target error uses every example error.
 
Stop when all Predictions are in range uses the difference between the actual desired output and the output predicted by the network.  This process can take a long time and the resulting network may not generalize well.
 
Stop when Average validating error is below the target error uses the total validating example errors divided by the number of validating example.
 
Stop when Average validating error is increasing.  Checks for increase in the average validating error for six cycle
 
6

Presentation

6. Presentation
In every training cycle the presentation of training examples to the neural network is normally in the forward direction from the first example training row to the last example training row.  If Balanced is checked the presentation is first in the forward direction and then in the reverse direction.  If Random is checked example training rows are presented at random.  If Grouped is pressed the example training rows are presented in groups.  Every row is presented at least once and possibly three times every training cycle.
7

OK button

7. OK buttonPress to accept all settings.
 
8

Cancel button

8. Cancel buttonPress to reject all settings.
 

Created with help of DrExplain