Abstract: Techniques are described for predicting future behavior of links in a network and generating dynamic thresholds for link metrics for use in path selection. In one example, a computing system receives historical values of a link metric for links of a network. The computing system executes a machine learning system which processes the historical values of the link metric to generate: (1) a predicted future value of the link metric for each link; and (2) a threshold for the link metric indicating whether the predicted future value for each link is anomalous. The computing system computes a path based on the predicted future values of the link metric and the threshold for the link metric. The computing system provisions the computed path, thereby enabling a network device to forward network traffic along the computed path.

Abstract: A device may generate a first polynomial and a second polynomial, and may generate, based on the first polynomial, a primary path from a first network device to a second network device via a first set of intermediate network devices. The device may generate, based on the second polynomial, a secondary path from the first network device to the second network device via a second set of intermediate network devices, and may assign a point of the first and second polynomials to the device, to each of the first set of intermediate network devices and of the second set of intermediate network devices. The device may cause the primary path to be provided from the first network device to the second network device, and may cause the secondary path to be provided from the first network device to the second network device.

Abstract: A device may receive a machine learning model, training data, and test data, and may perform a unit test on the machine learning model to generate unit test results. The device may perform regression tests on the machine learning model, with the training data and the test data, to calculate model scores, create graphs, determine inference delays, and identify missing points for the machine learning model. The device may perform scale and longevity tests on the machine learning model, with the training data and the test data, to identify additional missing points and calculate a resource utilization for the machine learning model. The device may update the machine learning model, to generate an updated machine learning model, based on the unit test results, the model scores, the graphs, the inference delays, the missing points, the additional missing points, or the resource utilization.

Abstract: In some implementations, a monitoring device may obtain a plurality of time-series data streams respectively associated with a plurality of resources. The monitoring device may generate, using a plurality of machine learning models and based on the plurality of time-series data streams, a plurality of sets of multi-step forecast values, wherein each set of multi-step forecast values is associated with the plurality of resources. The monitoring device may determine, based on the plurality of sets of multi-step forecast values, a set of particular multi-step forecast values associated with the plurality of resources. The monitoring device may cause, based on the set of particular multi-step forecast values, one or more actions to be performed. In some implementations, the monitoring device may determine, based on the plurality of time-series data streams and the plurality of sets of multi-step forecast values, that a correlation exists between a first resource and a second resource.

Abstract: Techniques are described for predicting future behavior of links in a network and generating dynamic thresholds for link metrics for use in path selection. In one example, a computing system receives historical values of a link metric for links of a network. The computing system executes a machine learning system which processes the historical values of the link metric to generate: (1) a predicted future value of the link metric for each link; and (2) a threshold for the link metric indicating whether the predicted future value for each link is anomalous. The computing system computes a path based on the predicted future values of the link metric and the threshold for the link metric. The computing system provisions the computed path, thereby enabling a network device to forward network traffic along the computed path.