Post-processing methods for delay embedding and feature scaling of reservoir computers.

Area of Science:

• Machine Learning
• Complex Systems

Background:

• Reservoir computing is a machine learning approach adept at complex time series prediction.
• Delay embedding and input data projection are crucial for accurate predictions in reservoir computing.

Purpose of the Study:

• To introduce novel post-processing methods for enhancing reservoir computing performance.
• To demonstrate the efficacy of these methods, particularly multi-random-timeshifting, in improving prediction accuracy and efficiency.

Main Methods:

• Developed post-processing techniques that train on past node states at uniform or random time shifts.
• Introduced the multi-random-timeshifting method for randomly recalling previous reservoir node states.
• Validated these methods using readout data from a physical laser reservoir system.

Main Results:

• Post-processing methods improve reservoir computer prediction by increasing feature dimension and enhancing delay embedding.
• The multi-random-timeshifting method allows for smaller reservoirs with large feature dimensions.
• This method is computationally inexpensive to optimize and demonstrated effectiveness in a physical reservoir.

Conclusions:

• Post-processing, especially multi-random-timeshifting, offers a significant advancement in reservoir computing for time series prediction.
• These methods are practical for experimentalists and applicable to physical reservoir systems.
• The multi-random-timeshifting method presents a computationally efficient and effective approach for enhancing reservoir computing capabilities.