Intelligent publication generation method based on dynamic reward decomposition and cooperation of combined meta-learning

By constructing task dependency graphs and encoding causal paths using GNNs, and combining differentiable Shapley values ​​and variance control, the problems of credit transfer and reward allocation in complex tasks of multi-agent reinforcement learning are solved, achieving efficient and interpretable publication generation.

CN122287701APending Publication Date: 2026-06-26NAVAL UNIV OF ENG PLA
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NAVAL UNIV OF ENG PLA
Filing Date
2026-03-30
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

Existing multi-agent reinforcement learning (MARL) methods suffer from problems such as credit transfer not following causal paths, poor adaptability to dynamic dependencies, unstable subtask decomposition, unfair reward distribution, and insufficient interpretability in complex collaborative tasks, especially performing poorly in tasks such as military publication generation.

Method used

We adopt a dynamic reward decomposition and collaboration method based on combinatorial learning. By constructing a task dependency graph, we use GNN to encode topology and causal paths, dynamically generate subtask weights and individual reward allocations, and combine path masking, differentiable Shapley values ​​and variance control to form an integrated closed-loop design to ensure that reward decomposition and incentive alignment are aligned.

Benefits of technology

It achieves low-variance, fast convergence, adaptability to non-stationary topologies and stage transitions, and fair and interpretable credit allocation, improving the quality, efficiency, and robustness of publication generation, reducing rework rates and cold start costs, and enhancing generalization capabilities in highly coupled and non-stationary scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122287701A_ABST
    Figure CN122287701A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent publication generation method based on combinatorial learning, dynamic reward decomposition, and collaboration. The method includes: constructing and updating a task dependency graph (DAG) online; encoding topology and causal paths through a GNN reward decomposition meta-controller to generate path masks and individual reward weights; achieving fair alignment by combining differentiable Shapley values ​​and hierarchical critics; introducing a variance control mechanism to stabilize training; and jointly optimizing each module through end-to-end consistency training to output high-quality publications. The system includes corresponding functional modules, and the electronic equipment includes a processor and storage media. This invention solves the problems of inaccurate credit transfer, poor dynamic adaptability, and unfair reward allocation in existing technologies, achieving low-variance, fast convergence, high robustness, and interpretable fair allocation. It is suitable for automated collaborative generation of complex combinatorial documents such as military publications, academic journals, and technical manuals.
Need to check novelty before this filing date? Find Prior Art