T Cell–Inspired Algorithms - Immunological Computation: Theory and Applications - page 79

result (Figure 4.5d) that covers almost the entire self-space. h e parameter r , which
works as a threshold, controls the detection sensitivity. A smaller value of r gener-
ates more general detectors (i.e., covering a larger area) and decreases the detection
sensitivity. However, for a more complex self-set, changing the value of r from 8
(Figure 4.5b) to 7 (Figure 4.5c) generates a coverage with many holes in the nonself
area, and still with some portions of the self covered by detectors. h erefore, this
problem is not with the setting of the correct value for r , but a fundamental limita-
tion of the binary representation that is not capable of capturing the semantics of
the problem space. h e performance of the Hamming-based matching rules is even
worse; it produces a coverage that overlaps most of the self-space (Figure 4.5d).
In summary, advantages of string representation are (1) any data can be eventually
represented in binary form; (2) it is easy to analyze; and (3) it is good for textual
or categorical information. Its limitations include the comprehensibility problem
(di cult to interpret in the original problem space), the potential scalability issue
(string size and matching threshold value), and some di culty in combining with
other techniques (conventional algorithms, machine learning, etc.). Accordingly,
the representation and matching rule for an NSA needs to be chosen in such a way
that it accurately represents the data proximity in the problem space.

4.6 Real-Valued Negative Selection Algorithms

h e real-valued representation encodes each data item as a vector of real numbers,
where, the representation (self/nonself ) space, U corresponds to a subset of R n ;
only samples of one class are assumed. Specifi cally, such samples are considered to
be representative data from the self-space. h en, based on these samples, a model
of the self-set is built. For instance, the self-set can be considered to consist of all
points within a certain distance from each sample point.
Diff erent versions of real-valued negative selection (RNS) algorithms were pro-
posed so far; they include

A heuristic algorithm to generate “hyperspherical detectors”
NS with detection rules (an evolutionary algorithm to generate the “hyper-
cube detector”
Randomized RNS (an algorithm for generating hyperspherical detectors
using random process to optimize the distribution of detectors)
V-detector a lgorit hms
NS with fuzzy detection rules (an evolutionary algorithm to generate “fuzzy
rule detector”)

Moreover, RNS algorithms can also be classifi ed as (1) the “classical” generation-
and-elimination strategy (Gonzalez, 2003); (2) evolutionary approaches, for example,
GA (Dasgupta and Gonzalez, 2002); (3) one-shot randomized algorithm