On a Finitely Activated Terminal RNN Approach to Time-Variant Problem Solving

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Recurrent Neural Networks (RNNs) are crucial for time-variant computing.
• Asymptotic convergence in traditional RNNs presents limitations for dynamic problems.
• A need exists for RNN models with finite-time convergence and specialized activation functions.

Purpose of the Study:

• To develop and analyze terminal RNN models with finite-valued activation functions (AFs).
• To address the limitations of asymptotically convergent RNNs in time-variant computations.
• To demonstrate the efficacy of finite-time convergent RNNs for complex dynamic tasks.

Main Methods:

• Development of terminal RNN models incorporating finite-valued activation functions.
• Theoretical examination of finite-time convergence properties for neuron dynamics.
• Application of the proposed model to time-variant quadratic programming and manipulator motion planning.

Main Results:

• The proposed terminal RNNs exhibit finite-time convergence of error variables.
• The model effectively solves time-variant quadratic programming and redundant manipulator motion planning.
• Numerical results show convergence rates comparable to existing power-rate RNNs.

Conclusions:

• Finite-valued activation functions in terminal RNNs enhance performance in time-variant computations.
• The developed models offer a more desirable approach for dynamic problem-solving compared to asymptotic models.
• The proposed RNNs present a viable and effective solution for complex real-world dynamic systems.