A soft and hard constraint machine learning method and system based on abstract gradient descent

By employing an adversarial learning method based on abstract gradient descent, which alternately optimizes both hard and soft constraints, the balance between satisfying logical constraints and data fitting in machine learning models is addressed, achieving efficient soft constraint optimization under hard constraints.

CN119990237BActive Publication Date: 2025-10-21NAT UNIV OF DEFENSE TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411940369.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-12-26
Publication Date
2025-10-21
Estimated Expiration
2044-12-26

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively balance the demands of soft and hard constraints on machine learning models, especially in fields like autonomous driving, where they cannot guarantee that the learning model simultaneously satisfies both data fitting and logical rule constraints.

Method used

We adopt an abstract gradient descent-based approach, which uses adversarial learning to alternately optimize soft and hard constraints. By leveraging the properties of abstract gradients, we directly use a logistic semantic discrete function as the optimization objective. By combining backward and up-down abstraction strategies, we ensure that soft constraints are optimized within a feasible range after hard constraints are satisfied.

Benefits of technology

It achieves efficient optimization of soft constraints while satisfying hard constraints, avoids limited optimization space, and ensures that the learning model effectively fits the data under logical constraints.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119990237B_ABST
    Figure CN119990237B_ABST
Patent Text Reader

Abstract

The application discloses a soft and hard constraint machine learning method based on abstract gradient descent, and the method comprises the following steps: S1, establishing a learning model; S2, constructing a soft constraint from a training set, and constructing ∑ <x,y>∈D Loss(F α (x),y) loss function; S3, constructing a hard constraint from a logical property, and constructing min b F P (b,α); S4, performing an antagonistic alternating optimization between the soft and hard constraint optimization targets; and S5, outputting a trained model. The system is used for implementing the above method. The application has the advantages of simple principle, easy implementation, convenient operation, high efficiency and the like.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Method and device for constructing abstract model for target program

    CN118331846A

  • Systems and Methods for Multi-Objective Evolutionary Algorithms with Soft Constraints

    US20170169353A1