B Cell–Inspired Algorithms - Immunological Computation: Theory and Applications - page 132

5.2.3 AiNet and Its Variations

h is model is similar to RAIN and is proposed by De Castro and Von Zuben
(2001). h e main diff erence is that AiNet does not consider the stimulation con-
cept, rather uses the a nity concept. Part of the network adaptation process is
inspired by the clonal selection principle (Burnet, 1959). In AiNet, each network
element corresponds to an antibody molecule. h e a nity is used to remove redun-
dant information from the network—if the a nity between two antibodies is
greater than the suppression threshold, one of them is removed from the network.
Figure 5.7 shows the AiNet algorithm.
h e number of clones generated for one B cell (denoted as N c ) in the presence
of an antigen is computed as

N

∑

Nround Nd N
c

( )

(5.16)

ij

i 1

with N as the number of B cells in the population, d ij the distance between the
i th B cell and j th antigen, and round (

⋅

) is used to round a value to its closest

integer.

5.2.3.1 Opt-aiNet

De Castro and Timmis (2002a,b,c) uses a version of AiNet for solving optimiza-
tion problems, which is called opt-aiNet. h is work assumes B cell clusters as an
optimization problem (De Castro and Von Zuben, 2002a), where the center of
a cluster corresponds to a local optimum of the fi tness function (De Castro and
Timmis, 2002a). h us, clusters are expected to form around points with high fi tness
values.
In opt-aiNet (De Castro, 2003), B cells are encoded as real-valued vectors in
a Euclidean space. Also, a fi tness function to evaluate each B cell is defi ned based
on an objective function to be optimized (either minimized or maximized).
A population of B cells, considered as candidate solutions to the function being
optimized, evolves in AiNet. Such B cells undergo a process of evaluation against
the objective function, clonal expansion, mutation, selection, and evaluation of
their a nities with other B cells in the population. Accordingly, opt-aiNet fi nds
a B cell memory set that represents good values of the objective function. h e
network trains until it reaches a stable state, measured through the average fi tness
of the B cells (Figure 5.8).
If c is the cell to be mutated, then the resulting mutated cell c

′

(after a nity

proportional mutation) is computed as

′ = +

c c αε (5.17)