Instance-Specific Parameter Mapping for Combinatorial Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for solving combinatorial optimization problems using generic parameter values for classified problem data sets often result in insufficient solving performance.
Innovation Solution
A data processing system that classifies instances into classes based on feature vectors, calculates class centroid coordinates, determines axial directions for a class coordinate system, extracts instance groups, and generates unit vectors to correct generic parameter values, using a trained model to output appropriate parameter values for individual instances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If generic parameter values are used for classified problem data sets, then the solving process is simplified and faster, but the solving performance becomes insufficient
Solution Approach 1:
The patent applies local quality by transitioning from generic parameter values applied uniformly across entire classes to instance-specific parameter values tailored to each individual problem instance. The system calculates specific parameter values for each instance based on its features, allowing each instance to receive optimized parameters rather than generic class-level parameters, thereby improving solving performance while maintaining efficiency.
Solution Approach 2:
The patent implements parameter changes by dynamically calculating optimal parameter values for each problem instance based on its feature vector. Instead of using fixed generic parameters for entire classes, the system adjusts parameters individually for each instance through optimization calculations, enabling parameter adaptation that improves solving performance without sacrificing the efficiency gains from classification.
2Reliability
If instance-specific parameter optimization is performed, then solving performance improves, but the complexity of the system increases
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing optimal parameter values for each problem instance during an offline phase. The system processes training instances beforehand to determine their optimal parameters, stores these results, and then retrieves them during online solving operations. This eliminates the need for real-time optimization calculations, reducing online system complexity while maintaining instance-specific optimization benefits.
Solution Approach 2:
The patent uses copying by creating a database of pre-computed optimal parameter values for each problem instance. Instead of performing complex optimization calculations during the solving phase, the system copies and retrieves previously determined optimal parameters from stored data structures, significantly simplifying the online system architecture while preserving the performance advantages of instance-specific parameters.
3Reliability
If parameter values are optimized for each class, then solving performance improves compared to universal parameters, but further optimization for individual instances requires more computational resources
Solution Approach 1:
The patent applies preliminary action by performing the computationally intensive parameter optimization during an offline training phase rather than during online solving operations. The system pre-processes training instances to determine optimal parameters, stores these results in databases, and then simply retrieves them during actual problem solving. This shifts computational energy consumption to an offline phase, making the online system energy-efficient while maintaining high solving performance through instance-specific parameters.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A processing unit determines each axial direction of a class coordinate system having an origin corresponding to a class centroid, based on feature vectors of instances in the same class, obtains, for each second instance group corresponding to the axial directions, a set of second generic values for parameters, generates unit vectors representing correction directions for a set of first generic values, based on a first feature vector corresponding to the second instance group, the class centroid coordinates, and the sets of first and second generic values, creates a trained model that receives an instance in the class and outputs coordinates in the class coordinate system, and calculates, for a new instance classified into the class, values for the parameters for solving the new instance, based on first coordinates of the new instance obtained using the trained model, the set of first generic values, and the unit vectors.