Experimentos de aprendizaje con Máquinas de Boltzmann de alto orden
- Graña Romay, Manuel
- D'Anjou D'Anjou, Alicia Emilia
- Albizuri Irigoyen, Francisco Xabier
- Lozano Alonso, José Antonio
- Larrañaga Múgica, Pedro
- Yurramendi Mendizabal, Yosu
- Hernández Gómez, María del Carmen
- Jiménez Jiménez, José Luis
- Torrealdea Folgado, Francisco Javier
- Poza, M.
- González Acuña, Ana Isabel
ISSN: 0214-932X
Year of publication: 1996
Volume: 29
Issue: 4
Pages: 42-57
Type: Article
More publications in: Informática y automática: revista de la Asociación Española de Informática y Automática
Abstract
This paper presents the results obtained applying High Order Bolzmann Machines without hidden units to a several well known problems that represent quite diverse learning paradigms. The learning algorithm of Boltzmann Machines easily generalises for machines discrete units (whose state spaces are discrete multivalued sets) and continuous units. The learning algorithm remains valid for continuous units not normalised in the [0,1] interval. Besides, the absence of hidden units allows the estimation of the mean activation levels (activation probabilities in the binary case) without using simulated annealing. This provides a substantial acceleration of the learning process, and a qualitative improvement of the results. The presentation, description and discussion of the experiments follows the historical order of our works.