Training Recurrent Neural Networks via Dynamical Trajectory-Based Optimization

TL;DR

This paper presents a novel trajectory-based optimization method for training recurrent neural networks, leveraging dynamical systems theory to find optimal solutions more effectively than traditional algorithms.

Contribution

The paper introduces a new optimization approach using projected and quotient gradient systems for training RNNs, with stability guarantees and improved performance over existing methods.

Findings

Outperforms genetic algorithms in training RNNs.

Provides better results than error backpropagation.

Ensures stability of solutions via Lyapunov theory.

Abstract

This paper introduces a new method to train recurrent neural networks using dynamical trajectory-based optimization. The optimization method utilizes a projected gradient system (PGS) and a quotient gradient system (QGS) to determine the feasible regions of an optimization problem and search the feasible regions for local minima. By exploring the feasible regions, local minima are identified and the local minimum with the lowest cost is chosen as the global minimum of the optimization problem. Lyapunov theory is used to prove the stability of the local minima and their stability in the presence of measurement errors. Numerical examples show that the new approach provides better results than genetic algorithm and error backpropagation (EBP) trained networks.