Browse Prior Art Database

NETWORK AUTO-CORRECT DRIVEN BY MACHINE LEARNING

IP.com Disclosure Number: IPCOM000248365D
Publication Date: 2016-Nov-21
Document File: 6 page(s) / 831K

Publishing Venue

The IP.com Prior Art Database

Related People

Joe Clarke: AUTHOR [+4]

Abstract

Presented herein are machine learning classification and "recommender" algorithms to construct a network policy over an existing configuration command-driven "brown field" network. Based on the network policy, the system can "auto-correct" configuration changes that are out-of-policy. The result is a network with fewer configuration mistakes and configuration-driven downtime, which is the major cause of unplanned downtime.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 30% of the total text.

Page 01 of 6

NETWORK AUTO-CORRECT DRIVEN BY MACHINE LEARNING

AUTHORS:

    
Joe Clarke
Carlos M. Pignataro
Ganesh Kondaveeti
Bharath Kumar Gubbala

CISCO SYSTEMS, INC.

ABSTRACT

    Presented herein are machine learning classification and "recommender" algorithms to construct a network policy over an existing configuration command-driven "brown field" network. Based on the network policy, the system can "auto-correct" configuration changes that are out-of-policy. The result is a network with fewer configuration mistakes and configuration-driven downtime, which is the major cause of unplanned downtime.

DETAILED DESCRIPTION

    Networks are often complex and interrelated, and configuration mishaps are one of the leading causes of network outages. As such, it is crucial to understand the intent or policy defining a given network domain, and, based on that policy, only allow changes that are within its scope.

    Methods provided herein may be applied to an existing network with configurations and changes made previously therein. The methods involve defining a high-level policy that governs the network. Future configuration changes are, in real- time, analyzed and a prediction is made as to whether or not the proposed changes are in policy or not. This prediction includes impact to adjacent nodes and network-wide impact from point configurations. If the proposed changes are deemed to be not in policy, then a suggestion is made, if possible, as to what might be intended based on the aforementioned policy. Figure 1 below illustrates a graphical overview of the methods.

Copyright 2016 Cisco Systems, Inc.

1


Page 02 of 6

Figure 1

    Specifically, machine learning classification is used to determine a baseline policy by examining configuration changes, where in the network they are applied, what the change is, and who made the change. The training of the algorithm is performed as part of the creation of this baseline. When a future change is made, it is tested against the training set and a prediction is made as to whether this change is in policy. If so, the change is allowed. If not, the change may be implemented or not. If implemented, the training set is refined. If not implemented, a suggestion (i.e., auto-correct) is made.

    From a machine learning standpoint, the classification can use any number of algorithms to determine whether commands are in or out of policy. For example, a Bayesian network can examine commands either in the current network configuration or as they are configured and determine whether the commands fall in or out of policy. In the case where a current network configuration is loaded, the classifier might implicitly mark all the commands as being in policy. After that, the system is kept in a learning mode, and as changes are made to the network from that point, any command that is removed can be implicitly marked as out of policy. Alternatively, commands can be fed to the classifier and state explicitly whether they are in or out of policy relative to...