Optical Network Optimizer Using Neural Attention for Routing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current network optimization techniques based on analytic models, such as Queuing theory, are limited in handling the complexity of Software-Defined Networks and cannot efficiently improve global performance metrics like network utilization or latency planning, leaving room for improvement in routing optimization, network planning, and rapid failure recovery.
Innovation Solution
An optical network optimizer and optimization method utilizing a neural network with an attentional mechanism to accurately estimate key performance indicators, adjusting auxiliary output values and performing inference to optimize network performance by leveraging graph neural networks and message-passing operations for routing optimization, network planning, and rapid failure recovery.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If analytic models (e.g., Queuing theory) are used for network optimization, then the approach is simple and manageable, but it cannot handle the huge complexity of Software-Defined Networks and fails to improve global performance metrics
Solution Approach 1:
The patent replaces traditional analytic models (Queuing theory, network calculus) with a neural network-based system. The neural network learns optimal routing decisions through training on network state data, substituting mathematical analytics with a data-driven machine learning approach that can capture complex nonlinear relationships in SDN environments.
Solution Approach 2:
The patent transforms the optimization problem by changing from fixed analytic formulas to adaptive neural network parameters. The neural network's weights and biases are trained to optimize performance metrics dynamically, allowing the system to adapt to changing network conditions rather than relying on static analytical solutions.
2Ease of manufacture
If traditional optimization approaches are used, then the implementation is straightforward, but they are limited to improving only global performance metrics like network utilization or worst-case latency
Solution Approach 1:
The patent segments the network optimization problem into multiple independent routing decisions that can be made at different network nodes. The neural network processes local network states and generates routing decisions for specific flows, enabling distributed optimization rather than centralized control of all traffic.
Solution Approach 2:
The patent introduces dynamic adaptability through the neural network's ability to learn and update routing strategies based on current network conditions. The system transitions from static optimization rules to dynamic decision-making that adapts to real-time changes in traffic patterns, network topology, and performance metrics.
3Use of energy by moving object
If analytic models are applied to network optimization, then the computational requirements are low, but the accuracy of performance estimation and routing optimization is insufficient
Solution Approach 1:
The patent performs preliminary training of the neural network offline using historical network data. This pre-computation phase captures complex performance relationships in the trained model weights, enabling fast and accurate real-time routing decisions without repeated heavy computations during network operation.
Solution Approach 2:
The patent creates a virtual copy of the network's performance characteristics through the neural network model. The trained network learns and replicates the complex relationships between network states and performance metrics, providing accurate estimates without requiring computationally intensive simulations during actual routing decisions.
Data Source
AI summary
An optical network optimization method is disclosed. The optimization method includes training a neural network, adjusting at least one of a plurality of auxiliary output values of a plurality of auxiliary neurons of the neural network, and performing inference with the neural network. A neural network and an attention mechanism are utilized to predict network performance key performance indicator(s) so as to achieve efficient routing optimization, network planning and fast failure recovery.