Recently, two new learning algoriths, called TABP and HTABP and based on the properties of terminal attractors, have been proposed. These algorithms were claimed to perform global optimization of the cost in finite time, provided that a null solution exists. In this paper, we prove that, unfortunately, there are no theoretical guarantees that a global solution will be reached, unless the learning process begins in the domain of attraction of the global minimum. When a local minimum basin is entered, quite random jumps in the weight space take place, that may led to cycles. Moreover, when approaching local minima, overflow errors may also occur that force the learning to stop. Finally, particular care must be taken in order to avoid numerical problems that may occur even when approaching the global minimum.
Bianchini, M., Gori, M., Maggini, M. (1994). Does Terminal Attractor Guarantee Global Convergence?. In Proceedings of ICANN ‘94 (pp.377-380). Springer-Verlag.
Does Terminal Attractor Guarantee Global Convergence?
BIANCHINI, MONICA;GORI, MARCO;MAGGINI, MARCO
1994-01-01
Abstract
Recently, two new learning algoriths, called TABP and HTABP and based on the properties of terminal attractors, have been proposed. These algorithms were claimed to perform global optimization of the cost in finite time, provided that a null solution exists. In this paper, we prove that, unfortunately, there are no theoretical guarantees that a global solution will be reached, unless the learning process begins in the domain of attraction of the global minimum. When a local minimum basin is entered, quite random jumps in the weight space take place, that may led to cycles. Moreover, when approaching local minima, overflow errors may also occur that force the learning to stop. Finally, particular care must be taken in order to avoid numerical problems that may occur even when approaching the global minimum.File | Dimensione | Formato | |
---|---|---|---|
ICANN94.pdf
non disponibili
Tipologia:
Post-print
Licenza:
NON PUBBLICO - Accesso privato/ristretto
Dimensione
250.64 kB
Formato
Adobe PDF
|
250.64 kB | Adobe PDF | Visualizza/Apri Richiedi una copia |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/11365/29168
Attenzione
Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo