Network Graph Decomposition Using Unconstrained Resource Nodes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current supply chain planning methods are inefficient, taking several hours or even days to generate feasible and optimal plans due to the complexity of dealing with millions of decision variables and constraints, often exceeding the available time window for daily updates in manufacturing and distribution.
Innovation Solution
The technique involves decomposing a network graph into independent sub-graphs that share unconstrained resources, using machine-learning models to predict underutilized resources and relax constraints, allowing for faster processing by subdividing the graph into smaller, independent sub-graphs that can be analyzed in parallel.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the network graph is decomposed into more independent sub-graphs, then the processing efficiency is improved, but the accuracy of resource constraint prediction may deteriorate
Solution Approach 1:
The patent divides the network graph into multiple independent sub-graphs based on resource constraints identified by machine learning models. This segmentation allows parallel processing of each sub-graph, significantly improving computational efficiency while maintaining solution quality by preserving critical resource dependencies within each sub-graph.
Solution Approach 2:
The patent dynamically adjusts the decomposition granularity by changing parameters such as the confidence threshold for constraint prediction and the size of sub-graph clusters. This allows optimization of the balance between processing efficiency and prediction accuracy based on specific problem characteristics and available computational resources.
2Measurement precision
If machine learning models are used to predict unconstrained resources, then the decomposition accuracy is improved, but the computational overhead increases
Solution Approach 1:
The patent performs machine learning-based resource constraint prediction in advance during an offline training phase, generating pre-computed constraint information that is stored and reused during online graph decomposition. This preliminary action separates the computationally intensive model training from the time-critical decomposition process, reducing real-time computational overhead.
Solution Approach 2:
The patent uses trained machine learning models to generate copies of constraint prediction results for different scenarios and time periods. These pre-generated prediction copies can be quickly retrieved and applied during graph decomposition without re-running the complex training process, significantly reducing computational overhead during execution.
Data Source
AI summary
A network may be organized to provide components. Components may be generated by combining other components (sub-components) together, and components may be provided by resources in the network. This network may be represented by a graph of nodes representing components and resources. In order to efficiently analyze this graph to generate a network plan, the graph may be subdivided into independent sub-graphs. Individual resources may be shared by individual sub-graphs and considered independent when those resources are underutilized or otherwise unconstrained. Models may be used to predict which resources are unconstrained and allow those resources to be shared by otherwise independent sub-graphs, thereby increasing the decomposition of the graph and improving the efficiency of the network plan analysis.