# US 2023/0145097 A1 · Autonomous traffic (self-driving) network with traffic classes and passive and active learning

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2023/0145097 A1
- Document ID: US20230145097A1
- Status: Published application
- Kind: A1
- Application number: 18/052,614
- Filed: 2022-11-04
- Published: 2023-05-11
- Earliest claimed: 2021-11-05
- Assignees: The Regents of the University of California
- Applicants: The Regents of the University of California
- Inventors: Mariam Kiran; Nicholas Buraglio; Scott Campbell
- Web page: https://patentgrep.com/patent/US20230145097A1

## CPC

- H04L 47/127: by using congestion prediction
- H04L 41/147: for predicting network behaviour
- H04L 41/16: using machine learning or artificial intelligence
- H04L 43/08: Monitoring or testing based on specific metrics, e.g. QoS, energy consumption or environmental parameters
- H04L 45/08: {Learning-based routing, e.g. using neural networks or artificial intelligence}
- H04L 45/70: {Routing based on monitoring results}
- H04L 47/2483: involving identification of individual flows
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)

## Abstract

A method and apparatus for dynamically routing traffic from a first network device to a second network device is described. In one implementation, the method includes sampling one or more qualified packets from an ingress stream using neural network classification criteria, appending a flow encapsulation header with the path information, providing the path information to the path computation engine at user approved time intervals, transmitting the packets to destination network device, and collecting real-time network telemetry data. Real-time network telemetry data is stored in a network health database. When a new network transfer is introduced, the system correctly identifies the characteristics of the transfer, in terms of size, throughput, jitter, latency, loss, and completion times. A corrective action is introduced that inputs the current network health information and transfer characteristics by generating corrective control rules to allow the transfer to be optimally routed (e.g., automatically) through the network.

## Claims

### Claim 1

An apparatus for autonomous network traffic management, comprising:
(a) a non-transitory medium storing instructions executable by one or more processors;
(b) wherein said instructions, when executed by at least one of the processors, perform steps comprising:
(i) acquiring network traffic profiles and network health data from a network;
(ii) predicting future network statistics based on the network traffic profiles and network health data;
(iii) providing network optimization objectives; and
(iv) optimizing network path routing over the network optimization objectives utilizing deep reinforcement learning using the predicted future network statistics; and
(v) implementing the optimized network path routing over the network for particular network traffic profiles.

### Claim 2 (depends on 1)

The non-transitory medium of claim 1, wherein said acquiring network traffic profiles and network health data instructions when executed by the processor further performs steps comprising:
(a) inputting a sampling time frame;
(b) collecting network information over the sampling time frame;
(c) outputting the network information into an historical database;
(d) wherein the historical database comprises:
(i) a source in the network;
(ii) a destination in the network;
(iii) a transfer size;
(iv) a transfer duration; and
(v) network health data; and
(e) predicting an optimal solution for source-destination pathways for a certain traffic class.

### Claim 3 (depends on 2)

The non-transitory medium of claim 2, wherein said traffic classes are selected from a group of traffic classes comprising:
(a) long-living flows that contain multi-TB to multi-PB data transfers lasting 1-200 hours;
(b) short-lived flows that contain short amounts of “bursty” data transfers lasting less than 60 seconds;
(c) low-latency traffic flows that require low latency and minimum packet loss;
(d) priority flows that can be prioritized over other flows in the network;
(e) high throughput flows that require bandwidth greater than 100 GB/s;
(f) intent-based flows that have particular conditions associated with them such as deadlines, forbidden paths and required network characteristics; and
(g) default flows with no requirements attached.

### Claim 4 (depends on 2)

The non-transitory medium of claim 2, wherein said network health data is selected from one or more of a group of data comprising:
(a) latency;
(b) loss;
(c) bandwidth; and
(d) jitter.

### Claim 5 (depends on 2)

The non-transitory medium of claim 2, wherein said optimizing network path routing when executed by the processor further performs steps comprising:
(a) training a simulator Graph Neural Network (GNN) to provide an optimal solution for source-destination pathways; and
(b) simulating the traffic classes on the GNN to output the optimal solution for source-destination pathways.

### Claim 6 (depends on 5)

The non-transitory medium of claim 5, wherein said implementing the optimized network path routing over the network for particular network traffic profiles when executed by the processor further performs steps comprising:
(a) communicating routing actions directly to network elements; and
(b) utilizing the routing actions to transfer a future data transfer from a designated source to a designated destination according to the optimal source-destination pathway for the traffic class of the future data transfer.

### Claim 7 (depends on 5)

The non-transitory medium of claim 5, wherein said simulator Graph Neural Network (GNN) is hardware accelerated.

### Claim 8 (depends on 7)

The non-transitory medium of claim 7, wherein said hardware acceleration is accomplished by utilization of a Raspberry Pi or a field programmable gate array (FPGA).

### Claim 9

An apparatus for autonomous network traffic management, comprising:
(a) one or more network traffic monitors;
(b) one or more processors configured to receive input from one or more of the network traffic monitors; and
(c) a non-transitory memory storing instructions executable by the processor;
(d) wherein said instructions, when executed by the processor, perform steps comprising:
(i) detecting network health on at least one of the network traffic monitors;
(ii) identifying network traffic over a period of time into traffic classes;
(iii) identifying network health over the period of time into network health data;
(iv) storing network traffic information into an historical database;
(v) providing the traffic classes information, network health data and network traffic information to at least one AI/ML (Artificial Intelligence/Machine Learning) library, wherein the AI/ML library comprises a graph neural network (GNN) component;
(vi) predicting future network health from the AI/ML library graph neural network (GNN) component to create a Predicted and Optimal network solution; and
(vii) outputting the Predicted and Optimal network solution to a Route Behavior Engine (RBE).

### Claim 10 (depends on 9)

The non-transitory medium of claim 9, wherein said identifying network traffic and network health data instructions when executed by the processor further performs steps comprising:
(a) inputting a sampling time frame;
(b) collecting network information over the sampling time frame;
(c) outputting the network information into an historical database;
(d) wherein the historical database comprises:
(i) a source in the network;
(ii) a destination in the network;
(iii) a transfer size;
(iv) a transfer duration; and
(v) network health data; and
(e) predicting an optimal solution for source-destination pathways for a certain traffic class.

### Claim 11 (depends on 10)

The non-transitory medium of claim 10, wherein said traffic classes are selected from a group of traffic classes comprising:
(a) long-living flows that contain multi-TB to multi-PB data transfers lasting 1-200 hours;
(b) short-lived flows that contain short amounts of “bursty” data transfers lasting less than 60 seconds;
(c) low-latency traffic flows that require low latency and minimum packet loss;
(d) priority flows that can be prioritized over other flows in the network;
(e) high throughput flows that require bandwidth greater than 100 GB/s;
(f) intent-based flows that have particular conditions associated with them such as deadlines, forbidden paths and required network characteristics; and
(g) default flows with no requirements attached.

### Claim 12 (depends on 10)

The non-transitory medium of claim 10, wherein said network health data is selected from one or more of a group of data comprising:
(a) latency;
(b) loss;
(c) bandwidth; and
(d) jitter.

### Claim 13 (depends on 10)

The non-transitory medium of claim 10, wherein said optimizing network path routing when executed by the processor further performs steps comprising:
(a) training the Graph Neural Network (GNN) to provide an optimal solution for source-destination pathways; and
(b) simulating the traffic classes on the GNN to output the optimal solution for source-destination pathways.

### Claim 14 (depends on 13)

The non-transitory medium of claim 13, wherein said implementing the optimized network path routing over the network for particular network traffic profiles when executed by the processor further performs steps comprising:
(a) communicating routing actions directly to the Route Behavior Engine (RBE); and
(b) utilizing the routing actions to transfer a future data transfer from a designated source to a designated destination according to the optimal source-destination pathway for the traffic class of the future data transfer.

### Claim 15 (depends on 13)

The non-transitory medium of claim 13, wherein said simulator Graph Neural Network (GNN) is hardware accelerated.

### Claim 16 (depends on 15)

The non-transitory medium of claim 15, wherein said hardware acceleration is accomplished by utilization of a Raspberry Pi or a field programmable gate array (FPGA).

### Claim 17

An apparatus for autonomous network traffic management, comprising:
(a) a non-transitory medium storing instructions executable by one or more processors;
(b) wherein said instructions, when executed by at least one of the processors, perform steps comprising:
(i) providing one or more network optimization goals for a network;
(ii) acquiring network traffic profiles and network health data from the network;
(iii) predicting future network statistics on a pretrained hardware accelerated Graph Neural Network (GNN) based on the network traffic profiles and network health data;
(iv) inputting the future network statistics, network optimization goals, a network topology, and a network characteristic end point behavior into a Graph Optimizer to output optimized path information; and
(v) generating an optimized network path routing over the network using a Path Computation Engine (PCE), the optimized path information, and the network topology.

### Claim 18 (depends on 17)

The non-transitory medium of claim 17, wherein said future network statistics comprise one or more of: latency, jitter, utilization, and loss.

### Claim 19 (depends on 17)

The non-transitory medium of claim 17, wherein the GNN is periodically re-trained in an asynchronous manner based on network traffic characteristics.

### Claim 20 (depends on 17)

The non-transitory medium of claim 17, wherein the GNN is implemented in software or hardware selected from a set of hardware consisting of: a central processing unit (CPU), a graphical processing unit (GPU), and a field programmable gate array (FPGA).
