ALTERNATIVE ENERGY FUNCTIONS FOR OPTIMIZING NEURAL NETWORKS
Andrew H. Gee and Richard W. Prager
March 1992
When feedback neural networks are used to solve combinatorial optimization problems, their dynamics perform some sort of descent on a continuous energy function related to the objective of the discrete problem. For any particular discrete problem, there are generally a number of suitable continuous energy functions, and the performance of the network can be expected to depend heavily on the choice of such a function. In this paper, alternative energy functions are employed to modify the dynamics of the network in a predictable manner, and progress is made towards identifying which are well suited to the underlying discrete problems. This is based on a revealing study of a large database of solved problems, in which the optimal solutions are decomposed along the eigenvectors of the network's connection matrix. It is demonstrated that there is a strong correlation between the mean and variance of this decomposition and the ability of the network to find good solutions. A consequence of this is that there may be some problems which neural networks are not well adapted to solve, irrespective of the manner in which the problems are mapped onto the network for solution.
If you have difficulty viewing files that end '.gz',
which are gzip compressed, then you may be able to find
tools to uncompress them at the gzip
web site.
If you have difficulty viewing files that are in PostScript, (ending
'.ps' or '.ps.gz'), then you may be able to
find tools to view them at
the gsview
web site.
We have attempted to provide automatically generated PDF copies of documents for which only PostScript versions have previously been available. These are clearly marked in the database - due to the nature of the automatic conversion process, they are likely to be badly aliased when viewed at default resolution on screen by acroread.