Neural Prediction Agent

The AbleNeuralPredictionAgent provides a prediction or regression function. A back propagation neural network is used as the prediction model. To use the neural prediction agent, the user must specify the data source (and meta data in the form of a *.dfn file). Generating a neural prediction agent will:

  1. Create Import beans to be used for training network weights and testing network accuracy.
  2. Generate a time series filter bean if the Window is greater than 1, or the Horizon value is greater than 0.
  3. Scan through the source data and create the translate filters required to transform and scale the data for input to and output from the neural network.
  4. Configure the neural network architecture with the hidden unit numbers specified.
  5. Create the data connections for the data to flow through the beans.

Once the prediction agent is configured, the user can choose to train the network manually via Step/Cycle/Run/Halt buttons on the Agent Editor toolbar, or automatically via the Train button on the customizer dialog. Automatic training is controlled by specifying the maximum number of training passes allowed and a desired accuracy parameter.

After the network bean has been trained, its weights will be locked and it can be connected to other import beans to perform its prediction function.