We propose a procedure for designing the layout of a fixed fan-in neur
al network based on the pRAM model. Using minimisation of the relative
entropy between environment and output probability distribution laws
as a target, and Amari's learning rule as a strategy, we show that con
sideration of the conditional entropy of the output of one node, given
the candidate nodes that provide its inputs, leads to a locally optim
um choice of the the connections. The procedure is computationally fea
sible when certain preference criteria are used to control the mutual
relevance of the pRAM nodes. We give some simple numerical examples of
this procedure. (C) 1997 Elsevier Science Ltd. All rights reserved.