IEEE Computational Intelligence Magazine - May 2022 - 78

matches or exceeds OAMseq in 12 of 14
cases and loses two times to OAMseq on
vowel and poker datasets. EMTAUC
matches or exceeds OAMseq in 11 of
14 cases and loses three times to OAMgra
on splice, vowel, and poker datasets. This
phenomenon illustrates the advantages
of EMTAUC.
In addition, Table V shows the comparison
results of EMTAUC against
MOEAs in terms of AUC. The datasets
used in this experiment can be found
from the LIBSVM and UCI web site
[5]. EMTAUC outperforms ETriCM
and MKnEA-AUC in 9 out of 12 cases,
ties once, and loses two times, which
further shows that our proposal has
excellent performance.
C. Effectiveness of Dynamic
Adjustment Strategy of AUCC
Our proposal establishes a multitasking
AUC optimization environment to take
advantage of the similarity across the
constructed cheap task AUCc and the
original task AUCE. However, the knowledge
carried by the task AUCc may be
incomplete. Therefore, a dynamic adjustTABLE
V The comparison of EMTAUC against MOEAs in terms of AUC.
DATASET
Australian
Transfusion
Sonarall
ETriCM
pima_ind ians_diabctes
Musk1
Musk2
Hill_Valley_without
Wla
Parkinsons
Spcctf
Splice
Fourclass
w/t/l
0.907(0.745)−
0.786(0.713)−
0.887(1.385)+
0.737(0.513)−
0.904(0.844)−
0.958(0.229) .
1.000(0.005)+
0.912(0.165)−
0.776(2.029)−
0.830(1.978)−
0.912(0.258)−
0.806(0.352)−
9/1/2
MKnEA-AUC
0.904(0.663)−
0.794(0.309)−
0.912(0.954)+
0.750(0.401)−
0.914(0.730)−
0.961(0.208) .
1.000(0.000)+
0.913(0.190)−
0.808(2.329)−
0.857(0.843)−
0.919(0.174)−
0.808(0.156)−
9/1/2
ment strategy of AUCc is proposed to
make full use of the knowledge carried
by the whole datasets. In this section, the
effectiveness of the dynamic adjustment
strategy of AUCc is analyzed.
Table VI shows the AUC of the
EMTAUC
0.927(0.0004)
0.832(0.0009)
0.856(0.010)
0.830(0.0004)
0.921(0.0005)
0.964(0.0004)
0.958(0.008)
0.969(0.0002)
0.892(0.004)
0.868(0.0008)
0.927(0.0001)
0.832(0.0003)
−
EMTAUC with and without the dynamic
adjustment strategy of AUCc (EMTAUC
and EMTAUC-I) on all datasets, where
the basic MTO solvers in EMTAUC are
respectively set as MFEAII, SBGA, and
EMEA. EMTAUC can be observed to
win EMTAUC-I on problems 3, 4, 8, 10,
and 15 in terms of AUC values, which
further illustrates the effectiveness of the
dynamic adjustment strategy of AUCc. As
shown in Fig. 3, the average value of
rank (ordinal) correlations among fitness
landscapes in these problems are relatively
low. By dynamically adjusting the
sampled data in the AUCc, new knowledge
is introduced into the multitasking
AUC environment to improve the performance
of EMTAUC.
D. Parameters Analysis
This section analyzes the effect of three
critical parameters in EMTAUC and
takes the algorithm EMTAUC-SBGA as
an example. Three key parameters are
summarized as follows: 1) ,d the generation
intervals for dynamic adjusting
strategy; 2) s, the ratio of the whole dataset
-0.195
-0.19
-0.185
-0.21
-0.205
-0.2
510152025303540
Generation Intervals for
Dynamic Replacement Strategy
(a)
-0.082
-0.08
-0.078
-0.076
-0.074
-0.072
510152025303540
Generation Intervals for
Dynamic Adjustment Strategy
(b)
FIGURE 5 The relationships between the generation intervals for dynamic adjusting strategy and the average AUC value obtained by EMTAUCSBGA
on two representative datasets, where the dotted line represents the average AUC value obtained by single-task GA, (a) diabetes, and
(b) australian.
78 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2022
log (AUC)
log (AUC)
IEEE Computational Intelligence Magazine - May 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2022
