Graph Model Segmentation for Memory-Constrained LM Solving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing number of independent variables in models leads to memory resource constraints, making it difficult to solve large models efficiently.
Innovation Solution
A model processing method that segments the graph structure of a model into smaller submodels with fewer independent variables, using a random segmentation approach and the Levenberg-Marquardt LM algorithm to update variable values, allowing parallel solving and efficient memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the entire model with a large quantity of independent variables is used as input for solving, then the model can be solved as a whole, but memory resources become insufficient to support the solving process
Solution Approach 1:
The patent divides the graph structure of the model into multiple subgraphs by randomly segmenting nodes (independent variables). Each subgraph contains a subset of independent variables, allowing the model to be solved in smaller chunks that fit within available memory resources. This segmentation enables the system to handle large-scale models that would otherwise exceed memory capacity.
2Quantity of substance
If the graph structure is segmented into smaller submodels, then memory resource consumption is reduced, but the solving process becomes more complex
Solution Approach 1:
The model is divided into multiple subgraphs through random node segmentation, creating manageable submodels that can be processed independently with lower memory requirements.
Solution Approach 2:
The patent employs a universal solving approach that handles both segmented subgraphs and their interconnections through a standardized process. The solving apparatus uses a consistent algorithm framework that can process any subgraph configuration, reducing the complexity burden despite the segmentation.
3Ease of manufacture
If random segmentation is used to divide the graph structure, then the segmentation process becomes easy to implement and efficient, but may create more conflicts between submodels
Solution Approach 1:
The graph structure is segmented by randomly dividing nodes into different subgraphs, which simplifies the segmentation implementation process while maintaining effectiveness in reducing memory usage.
Solution Approach 2:
The patent incorporates a conflict detection and resolution mechanism that identifies conflicts between submodels and adjusts the segmentation or solving process accordingly. This feedback loop ensures that random segmentation does not lead to excessive conflicts that would undermine the benefits of segmentation.
Data Source
AI summary
The present disclosure relates to model processing methods and devices. In an example method, a graph structure corresponding to a first model is obtained. The graph structure is randomly segmented to obtain at least two first sets, where each first set includes at least one first independent variable. A value of the first independent variable is updated by using a Levenberg-Marquardt (LM) algorithm. The random segmentation operation and the updating operation are repeatedly performed until a convergence condition of the first model is met to obtain solutions of a plurality of first independent variables in the first model.