Information Technology Reference
In-Depth Information
n
.
∑
=
aff
_
val
=
x
(
i
)
−
x
(
i
)
/
n
(3)
d
identified
d
i
1
The non-dominated Abs' affinity value is calculated as follows:
I.
if the size of
dominated Abs is not zero, the affinity value equals the minimum affinity value of
the dominated Ab divided by two;
II.
otherwise, the affinity value is calculated ac-
cording to Eqs. (4), where
N
is the size of non-dominated Abs.
.
N
n
∑∑
=
aff
_
val
=
(
x
(
i
)
−
x
(
i
)
/
n
)
/
N
(4)
nd
identified
j
j
1
i
=
1
In this way, Ag-Ab affinity is indirectly embedded in Abs' affinity since non-
dominated Abs always have the smallest affinity value (the highest affinity).
4.
Clonal Selection:
Clonal selection consists of three steps:
I.
Abs with the smallest
affinity value are selected, i.e. non-dominated Abs are always selected;
II.
Abs in
the remaining population with affinity value smaller than a threshold (
δ
) are se-
lected;
III.
unselected Abs are kept in a different set.
5.
Clone:
I.
for selected Abs, a maximum clone size (
N
cmax
) is pre-defined; then a
fraction of
N
cmax
is allocated to each selected Ab according to its affinity percentage,
i.e. the higher the percentage the larger the fraction is assigned;
II.
Unselected Abs
are cloned once regardless of their affinity.
6.
Affinity Maturation:
I.
selected Abs are submitted to
hypermutation
, i.e. one di-
mension of the Ab is randomly chosen to mutate; the mutation rate is proportional
to the affinity value (inversely proportional to affinity); the whole process is calcu-
lated using Eqs. (5).
II.
unselected Abs are submitted to
receptor editing
which
means that more than one dimension (two, in PAIA) are randomly chosen to mu-
tate; the mutation rate is calculated using Eqs. (5).
.
x
(
i
)
=
x
(
i
)
+
α
⋅
N
(
0
,
i
=
1
K
,
n
;
α
=
exp(
aff
_
val
)
/
exp(
1
(5)
new
old
where
N (0, 1)
is a Gaussian random variable with zero mean and standard devia-
tion 1.
i
represents the dimension that has been chosen to mutate.
7.
Reselection:
the mutated/edited clones and their corresponding parents are com-
bined together and reselected:
I.
all non-dominated Abs are selected;
II.
if the
number of current non-dominated Abs (NCR) is less than the initial population size
(IN), Abs from the next non-dominated front are selected according to their recal-
culated Abs' affinity value (the ones with smaller affinity values are favoured) to
fill the difference between these two; this process continues until the difference is
filled;
III.
only when NCR is greater than IN and greater than the number of the
non-dominated Abs in the last iteration (NPR) can
Network Suppression
be in-
voked to suppress any too-close Abs.
8.
Network Suppression:
the
Euclidian
distance in objective space between any two
Abs is calculated; if it is less than a predefined network threshold (
) the one with
the larger affinity value is suppressed and deleted; this operator is invoked in step 7
when certain conditions are satisfied.
9.
Iteration:
the process is repeated from step 2 until certain conditions are met.
σ
In the following, some differences between PAIA and previous research are high-
lighted. Further discussion can also be found in Section 5.