Systems and methods for using bottleneck structures to improve performance of machine learning-based modeling
A processor-implemented method for generating a digital model of a communications system using a bottleneck structure includes receiving information associated with a communications system including multiple elements. Each of the elements is configured to communicate with other elements of the communications system. A bottleneck structure is generated based on the information associated with the communications system. An artificial neural network (ANN) processes the bottleneck structure and the information associated with the communications system to generate a digital model corresponding to the communications system.
1. A processor-implemented method performed by one or more processors, comprising:
receiving information associated with a communications system including multiple elements, each of the multiple elements being configured to communicate with other elements of the communications system and the information associated with the communications system including topology information and flow information;
determining an information rate estimate based on the topology information and the flow information;
generating a bottleneck structure based on the information associated with the communications system, the bottleneck structure comprising a computational graph characterizing a state of the communications system with flows and links, the links offering communications resources with a limited capacity and the flows making use of the communications resources, the bottleneck structure revealing interactions of bottleneck links and system-wide ripple effects caused by perturbations in the communications system;
generating an embedding based on the information rate estimate and the bottleneck structure; and
processing, by an artificial neural network (ANN), the embedding over one or more convolutional layers to generate a digital model corresponding to the communications system.
2. The processor-implemented method of claim 1 , in which the ANN generates flow rate estimates and link bandwidth utilization.
3. The processor-implemented method of claim 2 , in which the information associated with the communications system comprises one or more of a flow latency or a packet drop probability.
4. The processor-implemented method of claim 1 , in which the ANN comprises a graph neural network (GNN).
5. The processor-implemented method of claim 1 , in which the ANN leverages information provided by the bottleneck structure to generate the digital model.
6. The processor-implemented method of claim 1 , in which the ANN leverages the bottleneck structure represented by a directed graph to implement message passing between nodes of the directed graph, wherein the nodes of the directed graph represent multiple elements of the communications system including the links and the flows.
7. The processor-implemented method of claim 1 , in which the communications system comprises one of a wireless telecommunications network, a datacenter network, a wide area network, an edge cloud, a local area network, a cluster of computing devices connected by a network, or a network-on-chip in a microprocessor.
8. An apparatus, comprising:
at least one memory; and
at least one processor coupled to the at least one memory, the at least one processor being configured to:
receive information associated with a communications system including multiple elements, each of the multiple elements being configured to communicate with other elements of the communications system and the information associated with a communications system including topology information and flow information;
determine an information rate estimate based on the topology information and the flow information;
generate a bottleneck structure based on the information associated with the communications system, the bottleneck structure comprising a computational graph characterizing a state of the communications system with flows and links, the links offering communications resources with a limited capacity and the flows making use of the communications resources, the bottleneck structure revealing interactions of bottleneck links and system-wide ripple effects caused by perturbations in the communications system;
generate an embedding based on the information rate estimate and the bottleneck structure; and
process, by an artificial neural network (ANN), the embedding over one or more convolutional layers of the ANN to generate a digital model corresponding to the communications system.
9. The apparatus of claim 8 , in which the ANN generates flow rate estimates and link bandwidth utilization.
10. The apparatus of claim 9 , in which the information associated with the communications system comprises one or more of a flow latency or a packet drop probability.
11. The apparatus of claim 8 , in which the ANN comprises a graph neural network (GNN).
12. The apparatus of claim 8 , in which the ANN leverages information provided by the bottleneck structure to generate the digital model.
13. The apparatus of claim 8 , in which the ANN leverages the bottleneck structure represented by a directed graph to implement message passing between nodes of the directed graph, wherein the nodes of the directed graph represent multiple elements of the communications system including the links and the flows.
14. The apparatus of claim 8 , in which the communications system comprises one of a wireless telecommunications network, a datacenter network, a wide area network, an edge cloud, a local area network, a cluster of computing devices connected by a network, or a network-on-chip in a microprocessor.
15. A non-transitory computer readable medium having encoded thereon program code, the program code being executed by one or more processors and comprising
program code to receive information associated with a communications system including multiple elements, each of the multiple elements being configured to communicate with other elements of the communications system and the information associated with the communications system including topology information and flow information;
program code to determine an information rate estimate based on the topology information and the flow information;
program code to generate a bottleneck structure based on the information associated with the communications system, the bottleneck structure comprising a computational graph characterizing a state of the communications system with flows and links, the links offering communications resources with a limited capacity and the flows making use of the communications resources, the bottleneck structure revealing interactions of bottleneck links and system-wide ripple effects caused by perturbations in the communications system;
program code to generate an embedding based on the information rate estimate and the bottleneck structure; and
program code to process, by an artificial neural network (ANN), the embedding over one or more convolutional layers of the ANN to generate a digital model corresponding to the communications system.
16. The non-transitory computer readable medium of claim 15 , in which the ANN generates flow rate estimates and link bandwidth utilization.
17. The non-transitory computer readable medium of claim 16 , in which the information associated with the communications system comprises one or more of a flow latency or a packet drop probability.
18. The non-transitory computer readable medium of claim 15 , in which the ANN comprises a graph neural network (GNN).
19. The non-transitory computer readable medium of claim 15 , in which the ANN leverages information provided by the bottleneck structure to generate the digital model.
20. The non-transitory computer readable medium of claim 15 , in which the ANN leverages the bottleneck structure represented by a directed graph to implement message passing between nodes of the directed graph, wherein the nodes of the directed graph represent multiple elements of the communications system including the links and the flows.
21. The non-transitory computer readable medium of claim 15 , in which the communications system comprises one of a wireless telecommunications network, a datacenter network, a wide area network, an edge cloud, a local area network, a cluster of computing devices connected by a network, or a network-on-chip in a microprocessor.
22. An apparatus, comprising:
means for receiving information associated with a communications system including multiple elements, each of the multiple elements being configured to communicate with other elements of the communications system and the information associated with the communications system including topology information and flow information;
means for determining an information rate estimate based on the topology information and the flow information;
means for generating a bottleneck structure based on the information associated with the communications system, the bottleneck structure comprising a computational graph characterizing a state of the communications system with flows and links, the links offering communications resources with a limited capacity and the flows making use of the communications resources, the bottleneck structure revealing interactions of bottleneck links and system-wide ripple effects caused by perturbations in the communications system;
means for generating an embedding based on the information rate estimate and the bottleneck structure; and
means for processing, by an artificial neural network (ANN), the embedding over one or more convolutional layers of the ANN to generate a digital model corresponding to the communications system.
23. The apparatus of claim 22 , in which the ANN generates flow rate estimates and link bandwidth utilization.
24. The apparatus of claim 23 , in which the information associated with the communications system comprises one or more of a flow latency or a packet drop probability.
25. The apparatus of claim 22 , in which the ANN leverages information provided by the bottleneck structure to generate the digital model.
26. The apparatus of claim 22 , in which the ANN leverages the bottleneck structure represented by a directed graph to implement message passing between nodes of the directed graph, wherein the nodes of the directed graph represent multiple elements of the communications system including the links and the flows.