Modify Training Settings Dialog

The Modify Training Settings Dialog allows you to set the values associated with training the prediction. It is also used for modifying the default settings for the training and optimization of predictions.

Ä    Note: Fields that are currently being used by the Solution Service cannot be modified.

&  For help with predictions, see Predicting and Modeling Financial Data.

Establishing Default Values

When this dialog is accessed from the Modify Options Dialog: Field Creation page, it can be used to modify the default values to use when creating new predictions. Three different sets of default values are available, which correspond to the default training settings available on the Predict a Value Wizard: Select Options Page.

·      Defaults for No Optimization

These defaults correspond to the option to train without optimizing the inputs (quick).

·      Defaults for Brief Optimization

These defaults correspond to the option to briefly optimize the inputs and settings.

·      Defaults for Full Optimization

These defaults correspond to the option to extensively optimize the inputs and settings (slow).

For defaults including optimization, the default optimization settings can also be modified by pressing the Optimization Settings… button, which will display the Training Optimization Settings Dialog.

&  For help with optimizing fields, see Optimizing Signals and Predictions.

Selecting Training Criteria

Neural networks typically learn by examining a set of data and adjusting a set of internal weights based on how far off the current model is from the desired results. This is a multi-pass process, requiring the neural network to examine the data multiple times to create a desirable model. Each pass that the neural network makes through the entire set of data is called an epoch.

¨    Target Samples-to-Weights Ratio

This value is used to specify the target value of the ratio between the number of data samples available for training and the number of weights in the neural network. This ratio is used by the default setting of allowing neural network to automatically maintain the best estimates for its values. It is also used as the basis for training optimization monitoring the number of weights in the neural network.

Ä    Note: When establishing defaults, this value is used for all groups of default settings.

Ä    Note: In TradingSolutions v2.1 and previous versions, this value was internally set to 10-to-1.

Ä    Note: This option is not available when using a Custom Solution Wizard DLL.

¨    Number of Training Passes

Neural networks begin the training phase with a random set of initial weights. Different sets of initial weights may provide better or worse initial conditions from which to find an optimal solution. To increase the probability that a desirable model will be found, multiple training passes can be run.

When multiple training passes are used, the training pass with the best cross validation results will be kept as the training for the neural network. If no cross validation set is being used, the training pass with the best training results will be kept.

¨    Number of Training Epochs

This value is used to set the maximum number of times all of the training data will be presented to the neural network during the training phase. Increasing this value allows the prediction a better opportunity to learn the data. However, it will also lengthen the training time and increase the likelihood of over-specialization on the training data if cross validation is not being used.

¨    Maximum Epochs w/o Improvement

This value is used to decrease the amount of time spent attempting to train the neural network when no improvement is being made. Normally, a neural network will train until the number of training epochs is reached, keeping the weights associated with the lowest error for the cross validation data. This set of weights is used since it is the best set of weights for the general case.

Typically, while a neural network is being trained, the cross validation error will improve with the training error up a certain point. After that, the cross validation error will increase while the neural network begins to over-specialize on the training data. Therefore, it is typically safe to stop training when the cross validation error stops improving after a given number of epochs.

þ    Include disabled inputs in neural network input allocation.

In TradingSolutions version 2.1 and earlier, inputs that were disabled as part of genetic optimization were still allocated as part of the neural network topology. While the information was not used for training, removing the inputs would change the initial conditions associated with the training process.

This checkbox indicates that any disabled inputs are included for the allocation of the neural network topology. It is typically only used to preserve the values of predictions created with older versions of TradingSolutions.

Ä    Note: This option is not available as a default value.

Ä    Note: This option is not available if optimizable functions are included as inputs to this prediction.

What Do I Do Next?

When you are done modifying the training settings, press the OK button. If you would prefer to exit this dialog without making modifications, press the Cancel button.

How Did I Get Here?

The Modify Training Settings Dialog is displayed for a single field when you press the Training Settings… button on the Predict a Value Wizard: Select Options page or the Modify Field Dialog: Training Settings page.

The Modify Training Settings Dialog is displayed for modifying default values when you press the Training Settings… button on the Modify Options Dialog: Field Creation page.