This correlation between the initial set of weights
and the quality of the solution resembles the existing
correlation between the initial antibody repertoire
and the quality of the response of natural immune
systems, that can be seen as a complex pattern
recognition device with the main goal of protecting
our body from malefic external invaders, called
antigens. Antibodies are the primary immune
elements that bind to antigens for their posterior
destruction by other cells [9]. The number of
antibodies contained in our immune system is
known to be much inferior to the number of possible
antigens, making the diversity and individual
binding capability the most important properties to
be exhibited by the antibody repertoire. In this
paper, we present a simulated annealing approach,
called SAND (Simulated ANnealing for Diversity),
that aims at generating a dedicated set of weights
that best covers the weight space, to be searched in
order to minimize the error surface. The strategy
assumes no a priori knowledge about the problem,
except for the assumption that the error surface has
multiple local optima. In this case, a good sampling
exploration of the error surface is necessary to
improve the chance of finding a promising region to
search for the solution. The algorithm induces
diversity in a population by maximizing an energy
function that takes into account the inverse of the
affinity among the antibodies. The weights of the
neural network will be associated with antibodies in
a way to be further elucidated.
