Radial Basis Function Neural Networks: Training, Outputs, and Persistence

Summary

The document outlines a classic radial basis function neural network with a hidden layer of radially symmetric neurons and a linear or sigmoid-style output layer. It describes choosing the output activation from the training target range: hyperbolic tangent for values between negative one and one, sigmoid for values from zero to one, and no activation for targets outside that range. The network is constructed with input and output dimensions and a maximum hidden-layer size; training determines how many hidden neurons are needed.

Training accepts flattened input and output patterns, an epoch limit, and an allowable error, and stops when either limit is reached. A calculation method returns network responses, while save and load methods persist topology, errors, and weights, rejecting incompatible topologies. XOR and integer arithmetic examples illustrate the class, but the document provides no financial dataset, forecasting evaluation, or evidence of trading usefulness. It is an implementation overview rather than a trading method.

Key ideas

• The described RBF network uses radially symmetric hidden units and a linear or nonlinear output layer.
• Output activation is selected according to the range of target values.
• Training is bounded by either a maximum number of cycles or an acceptable error.
• The actual hidden-layer size is learned subject to a configured maximum.
• The class supports computing responses and saving or loading network structure and weights, but gives no trading evaluation.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.