Scholarlink

<p>Echo State Networks (ESNs) have emerged as powerful tools for time series prediction, yet their performance heavily depends on reservoir structure, which traditionally relies on random weights independent of input data characteristics. This paper presents a theoretical framework and optimization approach for ESN reservoir design, demonstrating that reservoir weights should be adapted based on input data properties, and that both topology and weights significantly influence prediction accuracy. Building on theoretical insights about input-dependent reservoir behavior, we propose two complementary methods: a supervised approach that directly optimizes reservoir weights through gradient descent, and a semi-supervised technique that combines small-world and scale-free network properties with hyperparameter optimization. Our extensive experiments across multiple datasets, including synthetic chaotic systems (Mackey-Glass and NARMA time series) and real-world climate data, demonstrate that the proposed methods consistently outperform traditional ESNs by achieving substantially lower prediction errors. Most notably, our analysis reveals that edge connectivity parameters play a crucial role, second only to reservoir size in determining network performance. These findings provide important practical guidelines for ESN design and open new directions for automated reservoir optimization based on input data characteristics.</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

Input driven optimization of echo state network parameters for prediction on chaotic time series

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年9月26日

查看全文
  • Leila Gonbadi,
  • Habib Rostami,
  • Ebrahim Sahafizadeh,
  • Sanaz Rostami,
  • Mojtaba Mansouri Nejad,
  • Ahmad Shirzadi

摘要

Echo State Networks (ESNs) have emerged as powerful tools for time series prediction, yet their performance heavily depends on reservoir structure, which traditionally relies on random weights independent of input data characteristics. This paper presents a theoretical framework and optimization approach for ESN reservoir design, demonstrating that reservoir weights should be adapted based on input data properties, and that both topology and weights significantly influence prediction accuracy. Building on theoretical insights about input-dependent reservoir behavior, we propose two complementary methods: a supervised approach that directly optimizes reservoir weights through gradient descent, and a semi-supervised technique that combines small-world and scale-free network properties with hyperparameter optimization. Our extensive experiments across multiple datasets, including synthetic chaotic systems (Mackey-Glass and NARMA time series) and real-world climate data, demonstrate that the proposed methods consistently outperform traditional ESNs by achieving substantially lower prediction errors. Most notably, our analysis reveals that edge connectivity parameters play a crucial role, second only to reservoir size in determining network performance. These findings provide important practical guidelines for ESN design and open new directions for automated reservoir optimization based on input data characteristics.