NET-REJ:  Generate networks from the posterior by rejection sampling.

Net-rej generates networks from the posterior distribution given
training data using the architecture, priors, and data specifications
obtained from the given log file, to which the generated networks are
also written.  The rejection sampling procedure used is exceedingly
inefficient, and is useful only in providing a check (in very small
examples) on whether the Markov chain methods are working properly.

Usage:

    net-rej log-file max-index [ rejection-limit ]

The networks with indexes from zero up to the indicated index are
generated, as long as the number of rejections does not exceed the
limit (default is infinity).  If the log file already contains
networks with some of these indexes, only the networks with indexes
greater than the last one in the log file are generated.

Networks are generated from the posterior by sampling from the prior
(with an adjustment for noise variances in the case of real-valued
outputs), and then rejecting some generated states based on the
likelihood.

The observed acceptance rate is displayed on standard error at the end
of the program.

Two records are written to the log file for every network generated -
an 'S' record containing the hyperparameters ('sigma' values), and a
'W' record containing the parameters (weights, biases, and offsets).
An 'r' record recording the random number generator state is written
as well.

            Copyright (c) 1995 by Radford M. Neal