Information Technology Reference
In-Depth Information
interests become fitter, reproduce and proliferate, while those that do not re-
ceive positive feedback are eventually removed from the population. The elitist
character of GAs is reflected in the way selection for reproduction is performed.
In evolutionary IF systems chromosomes are commonly selected for reproduction
according to their relative fitness. This can be accomplished simply by selecting
a fixed percentage of the most fit chromosomes [6], but to control the pace of
evolution, this percentage may be varied according to the overall filtering perfor-
mance [8]. To more accurately mimic natural evolution, one may assign to each
chromosome reproduction probability proportional to its fitness using
roulette
wheel
selection [7]. In any case, the most fit individuals are more likely to mate
and produce offspring that inherit their features (keywords). Diversity is pro-
gressively lost as the optimum profile proliferates and takes over the population.
This can cause evolutionary IF systems to overspecialise to one of the topics of
interest and reach a state which inhibits further adaptation.
Fillipo Menczer proposes a remedy to this loss of diversity. Arachnid [1] and
InfoSpiders [9] are two similar systems that use a population of agents that
autonomously crawl the web and filter information on behalf of the user. To
avoid a bias towards the most successful individuals a local selection schema
is adopted. Individuals are not selected for reproduction by comparing their
fitness (e.g. by ranking them according to decreasing weight), but rather, to
reinforce diversity, each individual reproduces once its fitness is over a certain
threshold.
This solution points towards the direction of Artificial Immune Sustems (AIS).
AIS are not meant to be accurate models of the biological immune system, but
use relevant processes and concepts. Simply put
2
, the main actors of the im-
mune system are
antibody
molecules that are responsible for recognising a class
of foreign,
antigen
molecules. In the case of IF, antibodies typically correspond
to user profiles and antigens to information items. How well an antibody recog-
nisesanantigen,inotherwordstheir
a
nity
to the antigen depends on their
structure. In IF terms, anity usually corresponds to the relevance score that a
profile assigns to an information item. When the anity between an antibody
and an antigen is over a threshold the immune system's
primary
response is
triggered. The antibody clones rapidly and thus the concentration of successful
antibodies increases. The cloning process is not accurate, but is subjected to
somatic hypermutation
that results in slightly different antibodies, possibly a
better match to the invading antigen. Further diversity of antibody repertoire is
maintained through replacement of a percentage of antibodies with new types
of antibodies that the bone marrow produces. With these processes the immune
system achieves adaptive pattern matching in the presence of different types of
antigens.
At the same time the immune system should avoid recognising and destroying
the host organism's own molecules. This ability for self-non-self discrimination is
what makes the immune system a particularly appealing metaphor . According
to one view, it is achieved through
negative selection
, that causes immature
2
For more details in AIS see [10,11,12].
