Chaotic time series prediction using phase space reconstruction based conceptor network.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Time Series Analysis

Background:

• Reservoir computing (RC) frameworks like Conceptor networks offer efficient online learning capabilities.
• Traditional RC methods use fixed, randomly initialized reservoir connections, potentially limiting performance.
• The structure of the reservoir connection significantly impacts RC computational performance.

Purpose of the Study:

• To optimize the reservoir connection within Conceptor networks.
• To introduce a novel phase space reconstruction (PSR)-based reservoir generation method.
• To evaluate the effectiveness of the PSR-based method on time series prediction tasks.

Main Methods:

• Developed a phase space reconstruction (PSR)-based reservoir generation technique.
• Implemented and tested the PSR-based Conceptor network on time series prediction.
• Compared the PSR-based Conceptor against Conceptor networks with random and cortex-like reservoir topologies.

Main Results:

• The PSR-based reservoir generation method improved prediction accuracy in Conceptor networks.
• All tested Conceptor network topologies exhibited a nonlinear decline in accuracy with increased storage load.
• The PSR-based Conceptor demonstrated superior accuracy across various storage load conditions compared to other topologies.

Conclusions:

• Optimizing reservoir connections is crucial for enhancing Conceptor network performance.
• The proposed PSR-based reservoir generation method offers a significant improvement for time series prediction using Conceptor networks.
• The PSR-based approach provides robust accuracy under varying storage loads, highlighting its practical utility.