Real-World Applications - Immunological Computation: Theory and Applications - page 195

individual per antigen, and allowing many memory individuals to contribute to
antigen recognition are developed. h is immune approach is compared with a
genetic-based approach, and it showed to be consistently more e cient, generating
higher mutation scoring test sets at a less computational cost.

7.10 H y b r i d A p p r o a c h e s
7.10.1

Application in Neural Networks

In Dasgupta and Forrest (1999a, 1999b), ongoing work using simulated annealing
and immune metaphors applied to the problem of fi nding good initialization vec-
tor for neural networks is presented. h e strategy needs no prior knowledge about
the problem except the assumption that the error surface has multiple local optima.
But a good region to search needs to be sampled for a good solution. In simulated
annealing for diversity (SAND), an antibody is a possible solution for the weight
vector of a given neuron in a single layer of the network. Antigens (the training
data) are ignored here. h e goal is to maximize the Ab-Ab distance so that similar-
ity in the population is reduced. A nity is measured as the Euclidean proximity
between two points in an n -dimensional shape space. h e energy measure to be
optimized is the sum of the Euclidean distances among all vectors that represent
the Ab population. h ey stop the search process whenever the distribution of the
Ab population reaches a close-to-uniform distribution.

7.10.2 Applications in Genetic Algorithms

To address the issue of designing a GA with improved convergence characteristics,
particularly in the fi eld of design constraints, Slavov and Nikolaev (1998) proposed
a GA simulation of the immune system. GAs have been found to be very sensitive to
the choice of algorithm parameters when applied to design constraints. h e authors
used the idea of antibody-antigen binding. h e fi tness of a solution is not only
dependent on the objective function value and the design constraints, as it would be
in a traditional GA, but also on how well the solution matches the best solution. h e
algorithm then selects these better solutions to adapt, thus leading to a higher con-
vergence rate when compared to a traditional GA. Hajela et al. (1997) adopt a more
generic approach to the adaptive problem solving by the use of the immune net-
work metaphor such as B cells, T cells, macrophages, and the MHC. h e immune
algorithm is used to produce adaptive behaviors of agents. Hajela et al. also experi-
mented by removing the interaction of the T cell in the searching algorithm, and
present convincing results that the eff ect of the T cell on performance is signifi cant
as the solutions found with using the T cell result in lower-cost solutions overall.
Other similar applications of the immune network metaphor for multimodal
function optimization can be found in Toma et al. (2000), Fukuda et al. (1999),
and Mori et al. (1998). Here, the authors use somatic hypermutation and immune