Scholarlink

<p>In the domain of collaborative intelligent systems, ensuring the accuracy and reliability of decisions, especially under critical conditions, represents a formidable and ongoing challenge. This study introduces a novel verification technique aimed at improving the performance of multiple collaborative classifiers. We utilize a modal logic model to reason about the distributed information among classifiers and develop a specialized language to define verification properties. Our approach, the Multi-Agent System Knowledge-Sharing algorithm, evaluates these properties by checking the satisfaction of defined formulas within the model. One key challenge addressed is the difficulty of defining suitable properties. Despite this challenge, our empirical results reveal that even with suboptimal properties, the technique substantially enhances error reduction. Extensive evaluation on datasets such as Fashion-MNIST, MNIST, and Fruit-360 demonstrates significant improvements. For instance, using 1000 agents, the number of incorrect verifications in Fashion-MNIST decreased from 89 to 17; with 700 agents, incorrect verifications in MNIST dropped from 71 to 3; and with 80 agents, incorrect verifications in Fruit-360 were reduced from 233 to 0.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Meet MASKS: integrating distributed knowledge and verification for multi-agent systems

  • 期刊论文

  • 发表日期：

    2025年4月1日

查看全文
  • Majid Alizadeh,
  • Amirhoshang Hoseinpour Dehkordi,
  • Ali Movaghar

摘要

In the domain of collaborative intelligent systems, ensuring the accuracy and reliability of decisions, especially under critical conditions, represents a formidable and ongoing challenge. This study introduces a novel verification technique aimed at improving the performance of multiple collaborative classifiers. We utilize a modal logic model to reason about the distributed information among classifiers and develop a specialized language to define verification properties. Our approach, the Multi-Agent System Knowledge-Sharing algorithm, evaluates these properties by checking the satisfaction of defined formulas within the model. One key challenge addressed is the difficulty of defining suitable properties. Despite this challenge, our empirical results reveal that even with suboptimal properties, the technique substantially enhances error reduction. Extensive evaluation on datasets such as Fashion-MNIST, MNIST, and Fruit-360 demonstrates significant improvements. For instance, using 1000 agents, the number of incorrect verifications in Fashion-MNIST decreased from 89 to 17; with 700 agents, incorrect verifications in MNIST dropped from 71 to 3; and with 80 agents, incorrect verifications in Fruit-360 were reduced from 233 to 0.