Page 34 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

18

are created, each of which are modified by different mutations. Then, each modified
copy is regarded as one child. Asexual reproduction is employed with different muta-
chil-
dren obtained by applying one step of the Adam algorithm with different objective

functions.
In particular, in this work, we used the same mutation operations proposed in [5].
Moreover, in training , we not only ensure that the imputed values for missing com-
ponents ( m j 0) successfully fool the discriminator (as defined by the minimax game),

we also ensure that the values outputted by for observed components ( ) are
close to those actually observed. Therefore, a second loss term is added to objective
function of
- Minimax mutation:

(8)

- Heuristic mutation:

(9)

- Least-squares mutation:

(10)

Evaluation Two fitness function are used in the evaluation step. The first one computes
the quality of a generator, and the second one is used to measure the diversity. The
quality function is defined as:

(11)

and the diversity fitness score is defined as:
. (12)
Finally, the evaluation (or fitness) function of the proposed evolutionary algorithm is
given by:
(13)
where balances two measurements: generative quality and diversity. Overall, a rela-

tively high fitness score , leads to higher training efficiency and better generative
performance.

CITA 2023 ISBN: 978-604-80-8083-9