Time-shift selection for reservoir computing using a rank-revealing QR algorithm

Area of Science:

• Computational neuroscience
• Machine learning
• Nonlinear dynamics

Background:

• Reservoir computing (RC) is a powerful recurrent neural network approach for nonlinear system analysis.
• Incorporating time-shifts in reservoir signals has shown potential for improving RC performance.
• Optimal selection of these time-shifts remains a challenge.

Purpose of the Study:

• To develop a systematic, model-independent technique for selecting optimal time-shifts in reservoir computing.
• To enhance the performance accuracy of reservoir computers, particularly for analog hardware implementations.
• To demonstrate the efficacy of the proposed method across different reservoir computing architectures.

Main Methods:

• A novel time-shift selection technique is proposed, based on maximizing the rank of the reservoir matrix.
• A rank-revealing QR algorithm is employed to identify optimal time-shifts.
• The method is validated on an optoelectronic reservoir computer and a traditional recurrent neural network with tanh activation.

Main Results:

• The proposed technique effectively selects time-shifts by maximizing reservoir matrix rank.
• Improved accuracy was observed compared to random time-shift selection across tested reservoir computer types.
• The method's task-independent nature makes it broadly applicable, including to analog hardware.

Conclusions:

• The developed time-shift selection method offers a robust and effective way to enhance reservoir computing performance.
• This approach is directly applicable to physical reservoir computing systems without requiring system models.
• The findings suggest a significant advancement in optimizing reservoir computing for complex nonlinear tasks.