Machine Learning With Continuous Relaxation for Diverse Discrete Solutions

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing combinatorial optimization methods, such as those using Ising machines and continuous relaxation simulated annealing, struggle to efficiently find multiple solutions due to issues with penalty coefficients and local solutions, leading to inefficiencies and limited solution diversity.

Innovation Solution

A machine learning program that trains a model using a cost function where discrete variables are relaxed to continuous matrices, allowing for simultaneous optimization with multiple penalty coefficients and Bayesian optimization of these coefficients to obtain a plurality of solutions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If continuous relaxation solving method is used to search for optimum solution, then search efficiency is improved, but ability to obtain multiple solutions deteriorates

Engineering Contradiction:
Improvesearch efficiencyVSAvoidability to obtain multiple solutions
Core Design Contradiction:
ProductivityVSAdaptability or versatility

Solution Approach 1:

The patent segments the optimization process by introducing multiple penalty coefficients (λ1, λ2, ..., λk) that divide the search space into different regions. Each penalty coefficient guides the search toward different types of solutions, enabling the system to obtain multiple diverse solutions while maintaining search efficiency through the continuous relaxation framework.

Inventive Principle:
Principle #1Segmentation

2Manufacturing precision

If penalty coefficient method is used in combinatorial optimization, then solution accuracy is improved, but reliance on penalty coefficients increases and solution diversity decreases

Engineering Contradiction:
Improvesolution accuracyVSAvoidsolution diversity
Core Design Contradiction:
Manufacturing precisionVSAdaptability or versatility

Solution Approach 1:

The patent dynamically adjusts penalty coefficients during the training process. By varying penalty coefficients across different training iterations and using multiple penalty coefficients simultaneously, the system maintains solution accuracy while exploring diverse solution spaces, thus reducing reliance on any single penalty coefficient value and increasing solution diversity.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameter values of penalty coefficients during the optimization process. By using multiple different penalty coefficient values (λ1, λ2, ..., λk) and adjusting them dynamically, the system achieves both accurate solutions and diverse solution sets, overcoming the limitation of fixed penalty coefficient methods.

Inventive Principle:
Principle #35Parameter changes

3Device complexity

If discrete variables are relaxed to continuous matrix, then optimization process is simplified, but obtaining discrete solutions becomes difficult

Engineering Contradiction:
Improveoptimization process complexityVSAvoiddiscrete solution quality
Core Design Contradiction:
Device complexityVSManufacturing precision

Solution Approach 1:

The patent performs preliminary relaxation of discrete variables to continuous matrices to simplify the optimization process. Then, in a subsequent step, it applies discrete rounding techniques to convert the continuous solutions back to discrete solutions. This two-stage approach maintains the simplicity of continuous optimization while ensuring high-quality discrete solutions are obtained.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4579535A1Machine learning program, determination program, machine learning method, determination method, machine learning device, and determination device
Publication Date: 2025.07.02 FUJITSU LTD
  • EP4579535A1 patent drawingFigure 1
  • EP4579535A1 patent drawingFigure 2
  • EP4579535A1 patent drawingFigure 3A~3B

AI summary

A machine learning program for causing a computer to execute a process includes training a machine learning model by machine learning that uses a cost function in which each element of a matrix obtained by relaxing a discrete variable to be optimized to a continuous matrix becomes a discrete optimization problem as a cost function in a search process that performs a search by adopting continuous relaxation into the discrete optimization problem.