Function approximation on non-Euclidean spaces
Neural Networks
| Hi-index | 0.00 |
This article presents a new algorithm for the automatic generation of neural architectures and supervised learning. Given a set of examples, the algorithm generates an architecture and synaptic weights that are adapted to the sampled problem. The algorithm supports any type and amount of numerical input and output. The learning/generation process is guaranteed to converge in a strictly finite number of steps. With the exception of certain a priori nonoptimal architectures, all architectures without internal loops are potentially accessible, and the algorithm tends to generate architectures of minimal complexity, giving it high generalization performance in the learned domain.