Learning-Support Method for Professional Shogi Players

Using Emotions of Others

Takeru Isaka and Iwaki Toshima

NTT Digital Twin Computing Research Center, Tokyo, Japan

Keywords:

Personalized Learning, Collaborative Learning, Professional Player, Emotion, Japanese Chess.

Abstract:

Many methods have been proposed to support learning optimized for each learner. However, these methods

mainly target novice to intermediate learners. We propose a method for further improving the abilities of

advanced learners and professionals. At the advanced and higher levels, learners often target the behavior of

more competent others. However, it is difﬁcult to acquire the skills of such others simply by observing their

behaviors. A learner must understand the thought processes to arrive at their behavior. Knowing the emotions

that lead others to their behaviors could help learners understand others’ thought processes. On the basis

of this approach, we investigated a learning-support method that uses the emotions of others using Japanese

chess (Shogi) as the subject. We obtained valences, arousals, and subjective-position scores (i.e., evaluation

of whether black or white has an advantage for each position) for each move of Shogi from two professional

Shogi players with different playing styles. We observed noteworthy gaps in valence and arousal between

the two players, even with similar subjective-position scores. The players also gained new perspectives on

complex moves by referring to each other’s emotions. This suggests that awareness of the emotional gaps

with others can broaden a professional’s creativity.

1 INTRODUCTION

It is challenging for someone who has already reached

the top to further advance, regardless of the competi-

tion, art, or industry in which they excel (Yelle, 1979;

Argyris, 2002). Most learning-support methods are

targeted at novice to intermediate individuals. The ba-

sic way for advanced individuals to further improve

is through the learners’ self-help efforts. We pro-

pose a learning-support method using the emotions of

others to strengthen professional players of Japanese

chess (Shogi).

Shogi is a popular board game that is also played

professionally, and learning methods have been ac-

tively researched (Ito, 2018; Nishihara et al., 2018).

Novices can steadily improve their abilities through

manual learning methods such as memorizing the

roles of pieces and mastering the standard moves. An

example of a standard ﬁrst move in Shogi is P76,

which means to move the pawn to position 76.

In learning at this level, the optimal goal for

a novice called the “zone of proximal develop-

ment (ZPD)” (Vygotsky et al., 2011) can be clariﬁed

as a language. Clarifying goals as language means

that goals can be described as speciﬁc actions to be

taken by the learner (Skinner, 1954). The ZPD is de-

ﬁned as a task in which a learner cannot perform alone

but with outside help. For example, a novice can un-

derstand the reason from a textbook: P76 is necessary

to prepare to attack the opponent’s position by open-

ing the corner. Numerous learning-support methods

for novices have been proposed to automatically pro-

vide manualized knowledge and an optimal goal (Sid-

diqui et al., 2022; Malaise and Signer, 2022; Dicheva

et al., 2015; Rooein et al., 2022).

A typical method of learning Shogi played by pro-

fessionals is to look back at the history of moves

played in past games (hereafter referred to as “game

record”) and discover and learn decisive moves.

Game records include moves played by the Shogi ar-

tiﬁcial intelligence (Shogi AI), which is more skilled

than human Shogi players (Silver et al., 2018). In this

learning, it is relatively easy to grasp the intention of

the moves of others who are as competent as a learner

or have a similar playing style because there is only a

slight difference between their thoughts. However, it

is difﬁcult to grasp the intention of the moves played

by others who are more skilled than a learner or have

a different playing style because the difference be-

tween their thoughts is critical. Decisive moves by

486

Isaka, T. and Toshima, I.

Learning-Support Method for Professional Shogi Players Using Emotions of Others.

DOI: 10.5220/0012605100003693

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Computer Supported Education (CSEDU 2024) - Volume 2, pages 486-494

ISBN: 978-989-758-697-2; ISSN: 2184-5026

Previous learning (Observation and discussion)

Learning through others' emotions (Proposed)

Umm..

It is difficult to understand

more competent other’s behavior

by observation alone.

 Annotate emotions of oneself

and others to each step.

There’s discrepancy

in step (t-2)

Why are you comfortable

in step (t-2)?

Difficult..

Step

Emotion

t-1

t-2t-3

Noteworthy

emotion gap

Decisive move (Step t)

Decisive move

 Step with noteworthy emotion

gap

can be triggered to deeper

understand decisive move.

Figure 1: Overview of proposed learning-support method.

top professionals result from intuition and are difﬁ-

cult to explain logically, even for the top profession-

als who chose those moves (Ito et al., 2005). In other

words, it is difﬁcult to verbalize a learner’s ZPD at the

professional stages. The detection and presentation of

professional tacit knowledge is also a common chal-

lenge not only in Shogi but also in a wide range of

ﬁelds (Eraut, 2000).

1.1 Approach

We deﬁne emotion as how humans think about a situ-

ation one is facing. In Shogi, the situation is position

and move. In Shogi and chess, technical terms have

already been constructed to describe strategies and

patterns of moves. However, they are only a verbal-

ization of the typical thinking patterns of many play-

ers. Players with different playing styles have differ-

ent ways of thinking about situations that cannot be

explained in technical terms, and the higher the level

of players, the more remarkable these differences be-

come. The tacit knowledge we want to make effective

use of is, for example, the way of thinking that pro-

fessional players express when they say, “My ﬁngers

resist.” Such words are not always uttered, and even

if they are, it is difﬁcult for others to understand their

true meaning.

Previous studies have found that professional

board-game players have speciﬁc brain-activity pat-

terns compared with amateurs while playing (Wan

et al., 2011; Tanaka, 2018). They efﬁciently search

for the best next move by processing the positions

and moves as a lumped spatiotemporal pattern (Ito

et al., 2005; Ito and Takano, 2015). Ito and Matsub-

ara (Ito, 2004) suggested that positive-negative emo-

tions of the player accompany Shogi moves. Dis-

cussing with others and using their different perspec-

tives is effective in problem-solving (Miyake, 1986;

Dunbar, 1995). However, in discussions based on ob-

servation of game records, players do not always ver-

balize unconscious emotions.

If a learner can detect the transition of others’

emotions leading up to a decisive move, and if they

can identify the point of divergence with their emo-

tions, they will be able to discuss and dig deeper into

the differences in the way of thinking about the situa-

tion. Through this process (Fig. 2), it would be effec-

tive for professional players to develop and expand

their own emotions by taking in others’ emotions to

ﬁll in the gaps they have been unable to perceive. This

learning cycle is an essential aspect of learning sup-

port, regardless of the level of the learner (Skinner,

1954; Berlyne, 1966). Fig. 1 shows an overview of

the proposed learning-support method.

To provide a proof of concept for the above

learning-support method, we collected the emotions

of two female professional Shogi players (belong-

ing to the ladies professional Shogi-player’s associa-

tion of Japan) during Shogi play using Russell’s core

affect (Russell, 1980). We then displayed the col-

lected emotions of the two players in the time se-

ries and let them discuss the feedback on their Shogi

play. Finally, we conducted a qualitative evaluation of

whether referring to the collected emotions of others

is effective in improving professionals’ abilities. The

contributions of this study are as follows.

• We propose a learning-support method for clarify-

ing professional player’s ZPD using the emotions

of others.

• This study is the ﬁrst proof-of-concept case to

test the effectiveness of presenting the emotional

differences between two professionals to enhance

their learning.

Clarification of

learning points

by referencing

others’ emotions

Discussion on

differences in

emotions

Figure 2: Learning cycle considered for this study.

Learning-Support Method for Professional Shogi Players Using Emotions of Others

487

2 RELATED WORK

2.1 Personalized Learning Optimization

Numerous studies have incorporated AI to provide

optimized learning support for individual learners.

An essential factor common to these studies is the

automatic setting of optimal goals, referred to as

the ZPD. Several studies (Zou et al., 2019; Baker

et al., 2020) have shown that students’ performance in

school mathematics and English improves when they

select from tasks in the ZPD. Siddiqui et al. (Siddiqui

et al., 2022) proposed a method for university students

that uses keywords of interest to trigger the recom-

mendation of academic papers that the students are

curious about and to expand their knowledge. Malaise

and Signer (Malaise and Signer, 2022) proposed a

method for a learner of table tennis that automatically

suggests optimal exercises in the ZPD, taking into

account the learner’s knowledge and acquired skills.

Methods for improving interfaces to support learning

were also proposed using gamiﬁcation (Dicheva et al.,

2015) and chatbots (Rooein et al., 2022) to encour-

age students’ spontaneous learning for individualized

goals.

These studies targeted novice learners and stand

on the condition that their goals have been clearly

verbalized. It is difﬁcult to verbalize and conceptu-

alize the ideas and goals of professionals we focus

on as learning targets (Eraut, 2000). Although vague

and situational guidelines for learning to enhance pro-

fessionals have been discussed (Eraut, 1994; Argyris,

2002), they are far from being protocolized for imple-

mentation as learning-support methods. The learning-

support methods described above need to be extended

to cover a broader range of learner levels and do-

mains.

2.2 Game-Learning Support

Board-game learning support is a popular area of re-

search. Methods have been proposed to assist Shogi

novices by presenting predictive positions (Ito, 2018)

and visualizing battleﬁelds (Nishihara et al., 2018).

There are also learning-support methods for recom-

mending tasks tailored to the knowledge and abili-

ties of individual chess novices on the basis of the

ZPD (Guid et al., 2013).

Research has been conducted on chess (not Shogi)

to explain the difference in ability between human

experts and AI (P

alsson and Bj

ornsson, 2023; Schut

et al., 2023). However, to the best of our knowledge,

no personalized learning-support methods for profes-

sional board-game players have been reported.

2.3 Collection of Emotions

Certain emotion-collection approaches use biometric

data such as brain waves (Shen et al., 2009; Wan et al.,

2011; Nakatani and Yamaguchi, 2014). However, it is

difﬁcult to decipher the emotion expressed from bio-

metric details. We acquire emotions as annotations to

words that can be obtained in and under various ex-

perimental environments and conditions that are easy

for humans to interpret. As described in Section 3, we

set representative words (comfortable - uncomfort-

able, dynamic - gentle) that express the emotion of the

situation in playing Shogi and collected annotations

of closeness degree between the words and players’

emotions. The general purpose of collecting word an-

notations from experimental participants is to evalu-

ate average human emotions (Plutchik, 1980; Scherer

and Wallbott, 1994). In contrast, we focused on in-

dividual differences and aimed to record individual

emotions in a format others could reference.

2.4 Using Emotion in Learning

Many researchers have explored the integration of

emotions into learning-support methods. In school

education, it is known that understanding students’

emotional states has a meaningful impact on teach-

ers (Shen et al., 2009; Tyng et al., 2017), and methods

for communicating students’ emotions in the class-

room have been proposed (Sadiq. and Marentakis.,

2023). However, few researchers have examined

whether emotions effectively support learning among

professional-level learners.

Professionals’ tacit knowledge and intuition are

difﬁcult to verbalize, even for them both in Shogi (Ito

et al., 2005) and other ﬁelds (Eraut, 2000). It has

been suggested that tacit knowledge and intuition of

Shogi professional players are encoded in their minds

in a form similar to emotions (Ito, 2004). We investi-

gated whether visualizing professionals’ tacit knowl-

edge and intuition via emotions can help them learn.

3 SELECTION OF ANNOTATION

INDICATORS

We interviewed 17 advanced and the 2 professional

Shogi players mentioned above to determine whether

referring to emotion labels annotated by others for

each move would help them understand decisive

moves better than not referring to the annotation la-

bels. All the advanced players held at least the rank of

shodan (1-dan) or higher. The annotation labels used

in this interview were Pulchick’s basic emotions (joy,

CSEDU 2024 - 16th International Conference on Computer Supported Education

488

trust, fear, surprise, sadness, disgust, anger, and antic-

ipation) (Plutchik, 1980) and Russell’s core affect (va-

lence, arousal) (Russell, 1980), which are commonly

used emotion indices in the context of learning sup-

port. The interviews showed that 100% of the players

indicated that referring to the emotion labels anno-

tated by others is beneﬁcial.

Next, we studied the two professional play-

ers (same as in the interviews) to determine which la-

bels used in the previous step were valid. They played

six Shogi games and annotated all the labels in seven

levels for each move. We compared the validity of the

annotated labels by interview and degree of variance.

Fig. 3 shows the variance of annotated values per label

obtained in all six games. The horizontal axis repre-

sents the annotation label, and the vertical axis rep-

resents the variance of the annotated value. Russell’s

core affects had relatively large variances for both la-

bels, suggesting that the players were able to identify

and annotate differences in emotions for each move.

The interview results are summarized as follows.

• Pulchick’s basic emotions make it difﬁcult for

players to determine which eight labels are the

most critical for each situation.

• The labels of basic emotions other than joy and

anticipation rarely occur while playing Shogi.

• Core affects are highly consistent with the emo-

tions while playing Shogi, making them easy to

annotate without much thought.

On the basis of the above ﬁndings between the two

annotation indices, we considered that annotations by

core affects were suitable for extracting the unspo-

ken emotions expressed during Shogi games and used

them as an annotation index in the experiment de-

scribed in Section 4. Reﬂecting on the interview re-

sults, we set the label names of valence and arousal to

“comfortable - uncomfortable” and “dynamic - gen-

tle,” respectively, which are words generally used in

Shogi (e.g., “That move feels dynamic and comfort-

able.”).

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Variance of annotation values

Annotation label

Figure 3: Variance of annotation values.

4 ANNOTATION METHOD AND

EXPERIMENT

Using the annotation labels determined in the pre-

vious section, we conducted another experiment to

annotate emotions to moves and positions in Shogi.

The playing style of Shogi is generally expressed on

two axes: strategy (offensive and defensive) and mo-

mentum (aggressive and passive). The same two pro-

fessional Shogi players, with opposite playing styles

shown in Table 1, were also annotators for this exper-

iment to extract differences in emotion. Hereafter, we

refer to them as annotators 1 and 2.

The annotators annotated 24 games, including a

mix of three patterns: matches between pair of the

annotators and the Shogi AI, past great matches (the

annotators did not know the result of these matches),

and matches between two Shogi AIs. In the ﬁrst

match pattern, the annotator pair doubles as black,

and they discuss and decide on a move by consen-

sus. In the other two match patterns, the annotators

do not play; they only annotate. Hereafter, regard-

less of whether the annotator is a player, we refer to

the player who makes the move ﬁrst as black and the

player who makes the move second as white. We used

the graphical interface for Shogi (ShogiGUI, 2022)

and the Shogi engine (Demura, 2017) as the Shogi

AI and set it to a higher skill level than the annota-

tors. Considering the results of the survey described

in Section 3, we deﬁned the following three annota-

tion labels that are likely to express the traits of human

emotion while playing Shogi.

• Valence (Comfortable - Uncomfortable): Subjec-

tive feeling of comfort or discomfort for a move

from black’s standpoint.

• Arousal (Dynamic - Gentle): Subjective feeling

of aggressiveness for a move from black’s stand-

point.

• Subjective-position score: Subjective degree of

advantage or disadvantage of black for a position.

Valence and arousal range from -1 to 1, where

a lower value indicates a higher degree of “uncom-

fortable” or “gentle”, while a larger value indicates

a higher degree of “comfortable” or “dynamic”. The

subjective-position score ranges from -2000 to 2000.

The upper and lower limits indicate a decisive win or

Table 1: Annotators’ playing styles. They have opposite

playing styles, both in strategy and momentum.

Strategy Momentum

Annotator 1 Defensive Aggressive

Annotator 2 Offensive Passive

Learning-Support Method for Professional Shogi Players Using Emotions of Others

489

Fix

Finish

Pause

Dynamic

Comfortable

Gentle

Uncomfortable

Set valence and

arousal here

Set subjective-

position score here

Number

of moves

Evaluation for positionEvaluation for move

Figure 4: Interface of annotation tool designed to allow

plotting indicators (Valence, Arousal, Subjective-position

score) as continuous values selected to communicate emo-

tions that are difﬁcult to verbalize.

loss in the Shogi AI (Demura, 2017) we used in the

experiment. Lower values indicate an advantage for

white, while higher values indicate an advantage for

black.

To enable the annotators to quickly and intuitively

annotate after a move is determined, we developed

an annotation tool with which an annotation can be

completed with simple mouse operations, as shown

in Fig. 4. We used a two-dimensional interface, which

has been reported to allow intuitive and unloaded in-

put (Sadiq. and Marentakis., 2023). Our tool has va-

lence on the horizontal axis, arousal on the vertical

axis, and a straight line next to them for subjective-

position score. Annotation is completed with three

clicks: clicking into the two-dimensional coordinates,

clicking on the subjective-position score, and clicking

on “Fix”.

Fig. 5 shows the annotation procedure for Shogi

situations using this annotation tool.

1. First, black or white makes a move.

2. Next, the annotators annotate this move with three

types of emotions using the annotation tool from

black’s standpoint, regardless of whether it is a

move of black or white.

These steps are repeated until the end of the game.

After annotating each game, the annotators referred to

the annotation results and had a feedback discussion

about the game.

(1) Black makes

move

-2000

2000

(2) Annotators take position of black

and annotate black’s move and position

(4) Annotators take position of black

and annotate white’s move and position

(3) White makes

move

Repeat steps

(1) to (4) until

end of game

2000

-2000

Figure 5: Annotation procedure.

5 RESULTS

The average number of moves played per game in all

24 games was 109.8. Of these, the average num-

ber of moves with different positive and negative

signs (hereafter, referred to as “opposite”) among the

annotators was 13.5 in valence and 9.30 in arousal

per game in all 24 games. We conducted a between-

subjects analysis of variance (ANOVA) on the 2634

annotated emotions of all 24 games, and the results

indicate a signiﬁcant difference in valence between

annotators (F(1, 2633) = 6.595, p < 0.0001 for va-

lence, F(1, 2633) = 1.151, p = 0.689 for arousal, and

F(1, 2633) = 1.290, p = 0.529 for subjective-position

score).

Fig. 6 shows two annotation results of games with

noteworthy differences between the two annotators.

A between-subjects ANOVA was conducted from the

results in Fig. 6 (a), and there was a signiﬁcant dif-

ference in valence and subjective-position score be-

tween the annotators (p = 0.00287 for valence and

p = 0.00684 for subjective-position score). Fig. 6 (b)

shows the results that had opposite valences in 25

moves.

5.1 Emotion-Referenced Learning

The subjective-position scores in Fig. 6 (a) between

annotators became reversed from around move 43 and

remained opposite until the end of the game. In this

game, white won, so the subjective-position score at

the end of the game was expected to be negative.

Therefore, annotator 1 who annotated the end of the

game in Fig. 6 (a) with positive subjective-position

scores greatly misread the situation. Annotator 1

should correct her thoughts to moves in this game.

For annotator 1 to improve her skill, it is essential

to detect which moves were decisive for the game’s

CSEDU 2024 - 16th International Conference on Computer Supported Education

490

Figure 6: (a): Results that had signiﬁcant differences among

annotators. (b): Results that had many opposite emotions.

Upper graph of each result shows valence, and lower shows

subjective-position score, with horizontal axis representing

number of moves.

misreading and modify her thoughts and criteria for

evaluating the moves.

Fig. 6 (a) shows opposite valences in moves 22 to

37 despite the fact that the subjective-position scores

of both annotators are on the same side of the graph.

In such moves, the annotators had different ideas of

the move’s effect on the game. Annotator 1 may gain

new perspectives in modifying her thoughts by hear-

ing from annotator 2 about the reasons for the oppo-

site emotions. In other words, these moves with oppo-

site emotions are annotator 1’s learning points based

on the ZPD.

Table 2 lists dialogue excerpts from the annota-

tors’ discussion about the game in Fig. 6 (b) refer-

ring to annotated emotions. The annotators discussed

move 75, which has an opposite valence. This move

is superﬁcially agreed upon, since the positive and

negative signs of subjective evaluation coincide. In

other words, it is a situation that should be explored in

depth, which the annotators would have missed with-

out the emotional annotation. Displaying the emo-

tional transition of themselves and others enables an-

notators to detect the emotional gap and dig deeper

into what annotator 1 evaluated and what annotator 2

did not evaluate. In Table 2, the discussion triggered

by the emotional gap clariﬁed what both sides per-

ceive as beneﬁts and what they consider risks. Thus,

showing emotion can lead to complementary and bet-

ter next moves when logic cannot be explained in

words. With the proposed learning-support method,

learners can proceed with their learning in the follow-

ing steps.

• Learners annotate moves with their emotions and

make a time-series list of emotions.

• Using a list of emotions, learners detect emotional

gaps represented by moves with opposite emo-

tions.

• Learners discuss and dig deeper into emotional

gaps, focusing on the differences in thinking about

the advantages and disadvantages. They can pro-

mote understanding to approach the emotions of

others that they do not have, which leads to im-

proved abilities.

By referring to a time-series list of emotions cre-

ated by players with different playing styles, a player

can actively learn and incorporate the perspectives

one lacks, such as the risks and beneﬁts of a move.

This corresponds to an active learning cycle corre-

sponding to the change from diffuse curiosity to par-

ticular curiosity (Berlyne, 1966).

6 EVALUATION

We obtained subjective evaluations from the annota-

tors to subjectively evaluate the effectiveness of the

proposed learning-support method regarding other’s

emotions by using a ten-point scale. The higher the

score, the more effective our method, and the lower

the score, the more effective the standard learning

method of looking back at game records without re-

ferring to emotion. The two annotators gave an aver-

age evaluation score of 9.5 (Table 3).

We then interviewed a male professional Shogi

player (an instructor of the annotators who did not

participate in the experiment) about whether our

Learning-Support Method for Professional Shogi Players Using Emotions of Others

491

Table 2: Examples of actual discussions by annotators. The

dialogues collected in this study are in Japanese and trans-

lated into English by one of the authors.

Annotator 2: Is move 75 of white that

much advantageous?

Annotator 1: This was a pattern in which

the attack from white was effective.

Annotator 2: Although white may feel

comfortable, white had not regained the

pieces lost during the attack. Thus, white

has yet to achieve any practical result.

Annotator 1: I predict white’s T37 (a

move of Shogi) will reach the king in time.

Annotator 2: I don’t have that predic-

tion, because black’s Silver General (a kind

of Shogi piece) is effective and black has

many pieces in hand.

method helped improve professional ability. He an-

swered, “Professional players ﬁght on a ﬁeld beyond

the resolution of language. This approach is effective

in further strengthening those at a professional level.

The stronger the player, the more effective the method

is.” These qualitative evaluations and the results of the

interviews in presented in Section 3 suggest that our

method is effective in improving the ability of pro-

fessional Shogi players. It will be necessary to have

more professional players participate in the experi-

ment and conduct a statistical and long-term evalu-

ation to determine if the proposed method improves

their performance.

Table 3: Results of subjective evaluation by annotators us-

ing ten-point scale. Maximum score is ten.

Evaluation of annotator 1 10

Evaluation of annotator 2 9

7 DISCUSSION

The annotated results of the game in Fig. 6 (b) have

a marked distribution difference between the valence

and arousal shown in Fig. 7. From interviews with

the annotators, this was a reasonable result represent-

ing the difference in aggressive and passive playing

styles. The instructor of the annotators then com-

mented on his interpretation of this distribution dif-

ference as “Aggressive players (e.g., annotator 1) fo-

cus on whether their attacks work (comfortable) or

not (uncomfortable) and are sensitive to valence. On

the other hand, passive players (e.g., annotator 2) are

sensitive to the momentum of their opponent’s at-

tacks. This result shows such an emotional differ-

ence.” Although the number of professional players

Figure 7: Distribution of variance and arousal between two

professionals with different playing styles.

who participated in the experiment in Section 4 was

small, these interviews supported that the annotated

emotions reﬂect the tendency of professional players’

thinking.

Whether the proposed method can be generalized

to tasks other than board games is debatable. The

proposed method is suitable for turn-based tasks with

an objective evaluation axis for actions taken. It is

impractical to directly apply the proposed method to

tasks with continuously changing situations or tasks

with many participants. These tasks require automa-

tion of emotion extraction. The choice of words to be

used as annotation indicators may need to be modiﬁed

depending on the task domain.

7.1 White Boxing of AI

-2000

2000

You feel very

comfortable on

this move, Why?

Looks like a move

often played in

offensive style.

(b) Collaborative translation of Shogi AI’s complex

moves triggered by emotional discrepancies.

(a) Annotating emotions

for Shogi AI moves.

Arousal

Valence

Number of moves

Emotion

Figure 8: Collaborative translation of Shogi-AI moves us-

ing emotional gaps.

With the increase in computational resources and

evolution of computational Shogi algorithms (Silver

et al., 2018), even top professional Shogi players have

started to study the Shogi AI as a teacher (Nikkei Inc.,

2022). However, the process leading up to the move

decided by the Shogi AI is a black box. The proposed

method can facilitate learners’ understanding of com-

CSEDU 2024 - 16th International Conference on Computer Supported Education

492

plex Shogi AI moves by piecing together the emotions

of those who understand one aspect of the move, as

shown in Fig. 8.

• Several players with different playing styles anno-

tate the moves decided by the Shogi AI with the

emotions (Fig. 8 (a)).

• If each player can somewhat understand the AI’s

moves, they would be able to collaboratively

translate its moves by connecting the emotions of

multiple players (Fig. 8 (b)).

7.2 Learning with Emotion Estimator

Shogi AI's

best moves

˜B53

Current position

one step ahead

two steps ahead three steps ahead

˝B55 ˜P*54

˝S54

˝P56

˜B64

˝P55

Actual played

moves

˝: Learner’s move

˜: Shogi AI’s move

Learner’s scores on

lower branch are

estimated using DNN

-78

191

-782

191

-518

-403

-62

-424

˛: Learner’s subjective-

position score

˛: Shogi AI’s position score

Figure 9: Learning Shogi AI moves with emotion estimator.

The learning-support methods described in Sec-

tion 5.1 can only handle moves that were actually

played. If we can estimate learner’s emotions, we can

ﬁnd learning points on the basis of the ZPD from a

wide range that includes moves that have not been

played, represented by the Shogi AI’s potential de-

cisive moves.

Therefore, we created an emotion estimator that

takes the Shogi situation (position and move) as in-

put and outputs the emotions of the annotator. The

data collected from annotator 1 described in Sec-

tion 4 (2036 sets of inputs: positions and moves, out-

put: emotions) were split into training, validation, and

test data in the ratio of 4:1:1 to train a fully connected

deep neural network (DNN) consisting of one input

layer, four hidden layers, and one output layer. We

set the number of epochs as 800. The amount of data

is not yet sufﬁcient and the accuracy of the estima-

tion is not high, so the following results are for ref-

erence only. The top of Fig. 9 shows the branching

by the moves selected by the annotators in the game

in Fig. 6 (b), and the bottom shows the branching of

the Shogi AI’s (Demura, 2017) best moves (not ac-

tually played). We plotted the Shogi AI’s position

score and annotator 1’s subjective-position score for

each branch. Subjective-position scores on the lower

branch were estimated with the DNN. For the best

move of the Shogi AI, the difference between the

Shogi AI’s score and the learner’s score is smaller

than the upper branch. The lower branching sug-

gests that the learner may have a similar mindset to

the Shogi AI. Thus, the emotion estimator may help

detect potentially optimal goals for the learner based

on the ZPD.

8 CONCLUSIONS

To expand and develop a professional’s ability, we

proposed a learning-support method for extracting the

emotions of others and incorporating other’s thoughts

triggered by emotional gaps into the thoughts of

learners. We conducted an experiment to extract emo-

tions about Shogi moves from two professional Shogi

players and observed noteworthy emotional gaps be-

tween the two players, even with similar subjective-

position scores. The players were then able to use

the discrepancy in the others’ emotions as a trigger

for feedback discussions on the games. The qualita-

tive evaluations suggested that referencing the emo-

tion of others contributes to improving professional

Shogi players’ performance.

It will be necessary to conduct statistical and long-

term evaluations to determine whether this method

can be improved. We also plan to design a more ap-

propriate interface using emotional information that

can further enhance learning effects, as described in

Sec. 7.2. The proposed method is not limited to Shogi

but can be widely applied to other ﬁelds where pro-

fessional tacit knowledge and intuition are difﬁcult to

verbalize.

REFERENCES

Argyris, C. (2002). Teaching smart people how to

learn. Reﬂections-society for organizational learin-

ing, 4(2):4–15.

Baker, R., Ma, W., Zhao, Y., Wang, S., and Ma, Z. (2020).

The results of implementing zone of proximal devel-

opment on learning outcomes. International Educa-

tional Data Mining Society.

Berlyne, D. E. (1966). Curiosity and exploration: Animals

spend much of their time seeking stimuli whose sig-

niﬁcance raises problems for psychology. Science,

153(3731):25–33.

Demura, Y. (2017). Gikou. https://github.com/

gikou-ofﬁcial/Gikou (accessed 2024-2-22).

Dicheva, D., Dichev, C., Agre, G., and Angelova, G. (2015).

Gamiﬁcation in education: A systematic mapping

study. Journal of educational technology & society,

18(3):75–88.

Learning-Support Method for Professional Shogi Players Using Emotions of Others

493

Dunbar, K. (1995). How scientists really reason: Scientiﬁc

reasoning in real-world laboratories. The nature of in-

sight, 18:365–395.

Eraut, M. (1994). Developing professional knowledge and

competence. Psychology Press.

Eraut, M. (2000). Non-formal learning and tacit knowledge

in professional work. British journal of educational

psychology, 70(1):113–136.

Guid, M., Mo

zina, M., Bohak, C., Sadikov, A., and Bratko,

I. (2013). Building an intelligent tutoring system

for chess endgames. In International Conference

on Computer Supported Education, volume 2, pages

263–266. SCITEPRESS.

Ito, T. (2004). The thought and cognition on the verbal data

of shogi experts. Technical report, IPSJ SIG-GI-12-2.

Ito, T. (2018). Game learning support system based on fu-

ture position. ICGA Journal, 40(4):450–459.

Ito, T., Matsubara, H., and Grimbergen, R. (2005). Chunk-

ing in shogi: new ﬁndings. In Advances in Computer

Games, pages 140–154. Springer.

Ito, T. and Takano, D. (2015). Changes in cognitive pro-

cesses and brain activity. ICGA Journal, 38(4):209–

223.

Malaise, Y. and Signer, B. (2022). Personalised learning en-

vironments based on knowledge graphs and the zone

of proximal development. In CSEDU (1), pages 199–

206.

Miyake, N. (1986). Constructive interaction and the it-

erative process of understanding. Cognitive science,

10(2):151–177.

Nakatani, H. and Yamaguchi, Y. (2014). Quick concur-

rent responses to global and local cognitive informa-

tion underlie intuitive understanding in board-game

experts. Scientiﬁc Reports, 4(1):5894.

Nikkei Inc. (2022). The real of Shogi AI that gave birth

to Sota Fujii. https://bizgate.nikkei.com/article/

DGXZQOLM1405Q014042022000000 (accessed

2024-2-22).

Nishihara, Y., Takayama, R., Hishida, K., and Yamanishi,

R. (2018). Support method for beginners understand-

ing shogi games by evaluating battleﬁelds and a king’s

danger degree by using an arrangement of komas.

Journal of Japan Society for Fuzzy Theory and Intel-

ligent Informatics, 30:796–803.

alsson, A. and Bj

ornsson, Y. (2023). Unveiling con-

cepts learned by a world-class chess-playing agent. In

Proceedings of the Thirty-Second International Joint

Conference on Artiﬁcial Intelligence, pages 4864–

4872.

Plutchik, R. (1980). A general psychoevolutionary theory of

emotion. In Theories of emotion, pages 3–33. Elsevier.

Rooein, D., Paolini, P., Pernici, B., et al. (2022). Ed-

ucational chatbots: A sustainable approach for cus-

tomizable conversations for education. In CSEDU (1),

pages 314–321.

Russell, J. A. (1980). A circumplex model of affect. Journal

of personality and social psychology, 39(6):1161.

Sadiq., A. and Marentakis., G. (2023). Communicating

emotions during lectures. In Proceedings of the 15th

International Conference on Computer Supported Ed-

ucation - Volume 2: CSEDU, pages 558–565. IN-

STICC, SciTePress.

Scherer, K. R. and Wallbott, H. G. (1994). Evidence for

universality and cultural variation of differential emo-

tion response patterning. Journal of personality and

social psychology, 66(2):310.

Schut, L., Tomasev, N., McGrath, T., Hassabis, D., Pa-

quet, U., and Kim, B. (2023). Bridging the human-

ai knowledge gap: Concept discovery and transfer in

alphazero. arXiv preprint arXiv:2310.16410.

Shen, L., Wang, M., and Shen, R. (2009). Affective e-

learning: Using “emotional” data to improve learning

in pervasive learning environment. Journal of Educa-

tional Technology & Society, 12(2):176–189.

ShogiGUI (2022). http://shogigui.siganus.com/ (accessed

2024-2-22).

Siddiqui, S., Maher, M. L., Najjar, N., Mohseni, M.,

and Grace, K. (2022). Personalized curiosity engine

(pique): A curiosity inspiring cognitive system for stu-

dent directed learning. In CSEDU (1), pages 17–28.

Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai,

M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D.,

Graepel, T., et al. (2018). A general reinforcement

learning algorithm that masters chess, shogi, and go

through self-play. Science, 362(6419):1140–1144.

Skinner, B. F. (1954). The science of learning and the art of

teaching. Cambridge, Mass, USA, 99:113.

Tanaka, K. (2018). Brain mechanisms of intuition in shogi

experts. Brain and Nerve= Shinkei Kenkyu no Shinpo,

70(6):607–615.

Tyng, C. M., Amin, H. U., Saad, M. N., and Malik, A. S.

(2017). The inﬂuences of emotion on learning and

memory. Frontiers in psychology, page 1454.

Vygotsky, L. et al. (2011). Interaction between learning

and development. Link

opings universitet.

Wan, X., Nakatani, H., Ueno, K., Asamizuya, T., Cheng,

K., and Tanaka, K. (2011). The neural basis of intu-

itive best next-move generation in board game experts.

Science, 331(6015):341–346.

Yelle, L. E. (1979). The learning curve: Historical re-

view and comprehensive survey. Decision sciences,

10(2):302–328.

Zou, X., Ma, W., Ma, Z., and Baker, R. S. (2019). To-

wards helping teachers select optimal content for stu-

dents. In Artiﬁcial Intelligence in Education: 20th

International Conference, AIED 2019, Chicago, IL,

USA, June 25-29, 2019, Proceedings, Part II 20, pages

413–417. Springer.

CSEDU 2024 - 16th International Conference on Computer Supported Education

494