IEEE Computational Intelligence Magazine - August 2020 - 71

TABLE VIII List of 17 HaiFong
Professional Players.
SYMBOL

ENGLISH
NAME

values of inter-z-diff, we obtain the value
0.45, which we can think of as the
remaining strength adjustment required
by inter-game SA for the program to fit
to a player's actual strength. Now, let us
consider the difference between the final
z and the initial z (1.5), indicated in the
row of z-diff. If we average these absolute
values of z-diff, we obtain the amount of
adjustment required only through intergame SA, which is about 1.08. In other
words, by using intra-game SA, we are
able to adjust to the player's actual strength
with a shorter distance, and follow up by
fine-tuning with inter-game SA.
During the inter-game SA process,
most players can detect the strength

The Elo rating estimation results of
the 12 professional players are shown in
Figure 13. Compared with SSA, the
DSA mechanism estimated each player's
rating with slightly less accuracy
(R 2 = 0.5986). However, with SSA, the
player had to choose the appropriate
difficulty level based on their personal
experience. Unlike SSA, DSA is able to
adjust the program's strength automatically to fit its opponent's strength, resulting in a win rate that is close to 50%
(47% + 57%).
Table IX lists the z value, intra-z after
intra-games and final-z after 15 games for
each player. Let the inter-z-diff be intra-z
minus final-z. By averaging these absolute

4
6
Rank (Kyu)

programs, in this experiment, one single version of the program with DSA
was used for all participants. The participants were not informed of the
adjustment mechanism.
In Figure 12, we show the MDSA
estimation process for the participants
P1, P8, and P16 (corresponding to the
players in Table VIII). The strength
relation of the three players is P12
P8 2 P16, according to the Go Rating
website [25]. The intra-game SA
mechanism was used for the first
three games, where we can observe a
rapid change to the value of z. At the
end of the first three games, the z values appear to cor respond to each
player's strengths (P1 2 P8 2 P16) in
order. For the remaining 12 games,
inter-game SA was used; in contrast, z
adjustment between the games are
more stable.

GORATING

8
10
12
14
16

P1

HSU,
HAO-HUNG

3,341

P2

CHEN, CHI-JUI

3,223

P3

CHIEN,
CHING-TING

3,202

P4

LIN, LI-HSIANG

3,181

P5

LAI, CHUN-FU

3,058

P6

LI, WEI

3,027

P7

HUANG, SHIHYUAN

3,023

P8

TSAI,
CHENG-WEI

3,004

3

P9

LU, I-CHUAN

3,003

2

P10

HSU, CHING-EN

3,0002

P11

LIN, YEN-CHENG

2,978

P12

YANG,
TZU-HSUAN

2,975

P13

LIN, CHIEH-HAN

2,965

P14

NIU, SHIH-TE

2,919

P15

CHANG,
CHIA-HUAN

2,9002

P16

PAI, HSIN-HUI

2,795

P17

YU, LI-CHUN

2,770

18
10

20
z=2

30

40

z=1

70

80

z=0

90 100
z = -1

FIGURE 11 FoxWQ kyu rank chart over a period of 100 games.

(z)

1
0
−1
−2
−3
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Game
P1

2
The player does not have a GoRating on the website.
The listed rating is an estimate given by his/her peers.

50 60
Game

P8

P16

FIGURE 12 MDSA strength estimation of selected players.

AUGUST 2020 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

71
IEEE Computational Intelligence Magazine - August 2020

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - August 2020
