Algorithms

System Identification

\\

Use the System Identification algorithm to model both non-linear and linear relationships. In a typical use case you predict a number of target fields from their past values as well as from the past and current values of other feature fields. The System Identification algorithm is powered by a multi-layered, fully-connected neural network. System Identification supports incremental fit.

Note: The System Identification algorithm only works with numeric field values.

Parameters

  • The wildcard character is supported within target and feature fields.

  • Use the dynamics parameter to specify the amount of lag to be used for each variable.

    • The dynamics parameter is required.
    • Must be a list of non-negative integers separated by hyphens.
    • The number of non-negative integers listed must equal the number of target variables plus the number of feature variables.
    • The non-negative integers align with target and feature variables based on the order in which they are written.
    • One dynamic value is matched with each wildcard and the same amount applies to all fields matched by that wildcard.

  • The conf_interval parameter specifies the confidence interval percentage for the prediction.

    • Value must be between 1 and 99.
    • A larger number means a greater tolerance for prediction uncertainty.
    • Default value is 95.
    • The conf_interval number used with the fit command does not need to be the same number used with the apply command.

  • The layers option specifies the number of hidden layers and their sizes in the neural network.

    • Must be a list of positive integers separated by hyphens.
    • Option defaults to 64-64 for two layers, each of a size of 64.

  • Use the epochs option to specify the number of iterations during training.

    • Must be a positive integer.
    • Default value for epochs is 500.

Syntax

| fit SystemIdentification <target-fields> from <feature-fields> dynamics=<int-int-...> [conf_interval=<int>] [layers=<int-int-...>] [epochs=<int>] [into <model-name>]

You can apply the saved model to new data with the apply command.

| apply <model-name> [conf_interval=<int>]

You can inspect the model learned by System Identification with the summary command.

| summary <model-name>

Syntax constraints

System identification cannot be used with K-fold cross validation.

Examples

The following example uses three lags of Expenses, two lags of HR1, two lags of HR2, and three lags of ERP.

| inputlookup app_usage.csv | fit SystemIdentification Expenses from HR1 HR2 ERP dynamics=3-2-2-3

The following example uses three lags of Expenses, two lags of all fields that starts with HR, and three lags of ERP.

| inputlookup app_usage.csv | fit SystemIdentification Expenses from HR* ERP dynamics=3-2-3

The following example uses a fully-connected neural network with three hidden layers, each with a layer size of 64. The total number of layers in the neural network is five and comprised of one input layer, three hidden layers, and one output layer.

| inputlookup app_usage.csv | fit SystemIdentification Expenses from HR1 HR2 ERP dynamics=3-1-2-3 layers=64-64-64

The following example uses System Identification on a test set.

This image shows the results from running the System Identification algorithm on a test set. The Visualizations tab of the Machine Learning Toolkit is displayed.

Local availability

Source

Adapted from the Splunk AI Toolkit 5.6.4 documentation at /en/splunk-cloud-platform/apply-machine-learning/use-ai-toolkit/5.6.4/algorithms-and-scoring-metrics-in-the-ai-toolkit/algorithms-in-the-ai-toolkit (section: regressor).

Press Cmd/Ctrl+K to focus search. Esc to close.

Type to search the portal.