Information Technology Reference
In-Depth Information
In order to begin to understand the extremely poor performance of random creation
compared to the evolving libraries, we recorded the number of detectors created by
each solution; that is the number of antibodies left after negative selection. Figure 2
shows a plot of the mean number of detectors in each generation for the different
algorithms, averaged over the 25 runs. This reveals that for all random creation
variants the vast majority of the potential antibodies produced are screened out by the
negative selection process, so only a very few detectors remain. It is also notable that
evolving 1 library produces far fewer detectors than evolving 2, 3 or 4.
Model
10 0 .0 0
GA- 1 Lib
G A- 2 Libs
Ga - 3 Libs
GA - 4 Libs
80.00
Rn d - 1 Lib
Rn d - 2 Lib s
Rn d - 3 Lib s
Rn d - 4 Lib s
60.00
40 .0 0
20.00
0.00
G eneration
Fig. 1.
Coverage of best performing individual over 2000 generations (x axis). Each result
shows the % antigens matched (y axis) by antibodies created from a varying number of
libraries. The results using random creation are shown for comparison. Values averaged over
25 runs.
Interestingly, when the highest number of detectors per generation is plotted the
GA with 4 libraries creates more detectors than the 2, 3 and 1 libraries (in that order)
and vastly more than the random methods. Since the mean and best fitness had
converged by this time, this indicates that convergence had occurred, but around a
very “brittle” region, so that random mutations were producing a few very poor
individuals in each generation. This would imply a very “rugged” structure for the
library-landscape, (low fitness-distance correlation). Intuitively, as the number of
libraries increases, so the combinatorial effects of changing one element of any library
become more dramatic: changing one gene in a 1-library system only affects one
detector, but if in a 4-library system it makes that fragment rcb-match a self protein it
will make 216 detectors auto-reactive. Clearly this merits further investigation.
Table 2 presents the summary data from these experiments; all differences are
significant except those between the random variants.
