Multilayer neural networks: an experimental evaluation of on-line training methods

摘要

Artificial neural networks (ANN) are inspired by the structure of biological neural networks and their ability to integrate knowledge and learning. In ANN training, the objective is to minimize the error over the training set. The most popular method for training these networks is back propagation, a gradient descent technique. Other non-linear optimization methods such as conjugate directions set or conjugate gradient have also been used for this purpose. Recently, metaheuristics such as simulated annealing, genetic algorithms or tabu search have been also adapted to this context. There are situations in which the necessary training data are being generated in real time and an extensive training is not possible. This "on-line" training arises in the context of optimizing a simulation. This paper presents extensive computational experiments to compare 12 "on-line" training methods over a collection of 45 functions from the literature within a short-term horizon. We propose a new method based on the tabu search methodology, which can compete in quality with the best previous approaches. Scope and purpose Artificial neural networks present a new paradigm for decision support that integrates knowledge and learning. They are inspired by biological neural systems where the nodes of the network represent the neurons and the arcs, the axons and dendrites. In recent years, there has been an increasing interest in ANN since they had proven very effectively in different contexts. In this paper we will focus on the prediction/estimation problem for a given function, where the input of the net is given by the values of the function variables and the output is the estimation of the function image. Specifically, we will consider the optimization problem that arises when training the net in the context of optimizing simulations (i.e. when the training time is limited). As far as we know, partial studies have been published, where a few training methods are compared over a limited set of instances. In this paper we present extensive computational experimentation of 12 different optimization methods over a set of 45 well-known functions.