Excitation-Oriented Forgetting Recursive Least Squares

To address the deficiencies in neural network learning performance caused by the use of stochastic gradient descent (SGD) and conventional recursive least squares (RLS) in real-time learning tasks based on radial basis function neural networks (RBFNNs), this paper introduces a novel parameter optimization algorithm for training neural networks, termed excitation-oriented forgetting recursive least squares (EOFRLS). The EOFRLS algorithm allocates the forgetting factor based on the persistent excitation (PE) level and error at each time point of the training trajectory, thereby enhancing the learning performance of the neural network. Theoretical analysis and function approximation simulations demonstrate that EOFRLS achieves higher learning accuracy and faster convergence rates compared to SGD, conventional RLS, and RLS with a forgetting factor.