Scholarlink

<p>Hybrid storage systems integrating on-chain and off-chain data offer scalable solutions for big data management. However, current methods rely on static rules that limit adaptability in dynamic environments, causing high retrieval delays and increased storage costs. This study proposes a Multi-Agent Deep Reinforcement Learning (MA-DRL) model to optimize hybrid storage systems. The model comprises a coordinator agent and specialized agents for data categorization, blockchain, IPFS, and cache management. These agents make adaptive decisions based on metadata and receive feedback from Hybrid Smart Contracts (HSC). Extensive simulations evaluate the system in terms of cumulative rewards, retrieval delay, and storage transaction costs. Results show the proposed model significantly outperforms Single-Agent (SA) models, reducing storage transaction cost and retrieval delay by approximately 96% compared to SA. The main contributions are a dynamic MA-DRL system model for hybrid storage and retrieval management, and a DRL-based coordinator agent that improves workload distribution among agents.</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

Adaptive hybrid storage and retrieval management in blockchain-IPFS systems: a multi-agent deep reinforcement learning approach

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Sahar Bahrampour,
  • Mohammad Reza Rasouli,
  • Mohammad Fathian

摘要

Hybrid storage systems integrating on-chain and off-chain data offer scalable solutions for big data management. However, current methods rely on static rules that limit adaptability in dynamic environments, causing high retrieval delays and increased storage costs. This study proposes a Multi-Agent Deep Reinforcement Learning (MA-DRL) model to optimize hybrid storage systems. The model comprises a coordinator agent and specialized agents for data categorization, blockchain, IPFS, and cache management. These agents make adaptive decisions based on metadata and receive feedback from Hybrid Smart Contracts (HSC). Extensive simulations evaluate the system in terms of cumulative rewards, retrieval delay, and storage transaction costs. Results show the proposed model significantly outperforms Single-Agent (SA) models, reducing storage transaction cost and retrieval delay by approximately 96% compared to SA. The main contributions are a dynamic MA-DRL system model for hybrid storage and retrieval management, and a DRL-based coordinator agent that improves workload distribution among agents.