Information Technology Reference
In-Depth Information
In PAIA, the initial population size can be any number (even 1). However, only an
optimal initial size can lead to the most efficient way of dealing with the problem.
Most previous research did not emulate Clonal Selection. In PAIA, it is emulated
to fully exploit the selected Abs so that they have more opportunities to be cloned
and mutated in the early iteration steps, which can speed up the convergence.
Most previous works used a fixed clone size for every Ab. In PAIA the clone size
is adaptively decided by the number of selected Abs and their affinities.
A new method (Eqs. (5)) is proposed to calculate the mutation rate, which ensures
that the mutation rate is at least 0.37 1 . The exploration ability is thus preserved
even when all Abs converge to a single (sub) optimum.
In PAIA, the population size is not fixed, but is finally controlled by
. The popu-
lation is regulated by network suppression so that any too-close Abs are sup-
pressed. The way to invoke network suppression is adaptive to the search process.
σ
3.2 The Generic AIS Framework
Although PAIA is a specific MOP algorithm, the main structure of the algorithm can
be extracted as a generic AIS framework for MOP solving, as shown in Fig. 1.
Random Initialization
Yes
End
Stop?
No
Activation
Clonal Selection
Clone
Affinity Maturation
Reselection
No
NCR>NPR
& NCR>IN?
Yes
Network Suppression
Memory Set
Newcomers
Next Population
Fig. 1. Generic AIS framework for MOP solving (NCR: the number of current non-dominated
Abs; NPR: the number of non-dominated Abs in the last iteration; IN: the initial Abs size)
Two kinds of activation are emulated, namely Ag-Ab activation and Ab-Ab activa-
tion, so that one obtains information from both the objective space (Ag-Ab affinity)
and the decision variables space (Abs' affinity) to select Abs. The Clonal Selection
and Clone prefer good Abs by giving them more chances to be cloned so that they
always dominate the whole population. Affinity Maturation increases the diversity of
the population so that more objective landscape can be explored. Reselection ensures
that good mutants are inserted into the memory set and bad Abs apoptosis . Network
1 If Abs are normalized, then aff_val is within 0~1; so
is within 0.37~1 according to Eqs. (5).
α
Search WWH ::




Custom Search