Learning Machine Clustering for Accelerated Network Data Sharing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Low Power and Lossy Networks (LLNs) face challenges such as lossy links, low bandwidth, and limited resources, making it difficult to implement effective routing, Quality of Service (QoS), security, and traffic engineering, and the use of traditional machine learning algorithms is inefficient for predicting network behavior due to the complexity and large number of nodes.

Innovation Solution

The technique involves determining sharable variables among Learning Machines (LMs) hosted on Field Area Routers (FARs), calculating correlation values, and forming clusters to enable the sharing of correlated variables, with a Distributed Learning Coordinator (DLC) dynamically computing clusters and adjusting the sharing rate based on network resources and correlation, facilitating accelerated learning by sharing information between LMs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional machine learning algorithms are used to predict network behavior in LLNs, then the algorithms can process network data, but the processing efficiency is insufficient due to the complexity and large number of nodes

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidnetwork complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent divides the network into multiple clusters, each managed by a Learning Machine (LM) that processes local network data independently. This segmentation allows parallel processing across multiple LMs, significantly improving processing efficiency while reducing the computational burden on any single node in the complex LLN environment

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent combines the processing capabilities of multiple Learning Machines by having them share learned parameters and models. The LMs are trained independently on local data and then exchange their learned parameters to collectively improve prediction accuracy, effectively merging their computational power to handle the complexity of large-scale LLNs

Inventive Principle:
Principle #5Merging (Combining)

2Measurement precision

If Learning Machines are deployed on multiple Field Area Routers to improve learning accuracy, then the learning performance increases, but the network resource consumption and complexity increase

Engineering Contradiction:
Improvelearning accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments the learning task by deploying independent Learning Machines on different Field Area Routers, allowing each LM to specialize in processing local network patterns. This distributed architecture improves learning accuracy through diverse local insights while keeping individual node complexity manageable

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The Learning Machines use universal parameter structures that can be applied across different network contexts. By learning and sharing generic network behavior patterns that are applicable across multiple routers, the system achieves high learning accuracy without requiring complex specialized models at each node

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Productivity

If Learning Machines share information across all Field Area Routers, then the learning speed accelerates, but the bandwidth consumption and network traffic increase

Engineering Contradiction:
Improvelearning speedVSAvoidbandwidth consumption
Core Design Contradiction:
ProductivityVSQuantity of substance

Solution Approach 1:

The system extracts only the essential learned parameters from each Learning Machine for sharing, rather than exchanging complete model states or raw data. This selective extraction accelerates learning convergence while minimizing the bandwidth consumption and network traffic overhead in LLNs

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The LMs perform partial information sharing by exchanging parameters periodically or selectively based on learning progress, rather than continuously. This partial action approach maintains accelerated learning speed while controlling network resource consumption within acceptable limits

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP2954464B1Accelerated learning by sharing information between multiple learning machines
Publication Date: 2022.11.30 CISCO TECHNOLOGY INC
  • EP2954464B1 patent drawingFigure 1
  • EP2954464B1 patent drawingFigure 2
  • EP2954464B1 patent drawingFigure 3

AI summary

In one embodiment, variables maintained by each of a plurality of Learning Machines (LMs) are determined. The LMs are hosted on a plurality of Field Area Routers (FARs) in a network, and the variables are sharable between the FARs. A plurality of correlation values defining a correlation between the variables is calculated. Then, a cluster of FARs is computed based on the plurality of correlation values, such that the clustered FARs are associated with correlated variables, and the cluster allows the clustered FARs to share their respective variables.