Improving Active Attitude for Interactive Decision-making with Multiple
Agents by Increasing Personal Resource
Yoshimasa Ohmoto
1
, Masato Kuno
2
and Toyoaki Nishida
2
1
Department of Infomatics, Shizuoka University, Hamamatsu-shi, Shizuoka-ken, Japan
2
Graduate School of Infomatics, Kyoto University, Kyoto, Kyoto, Japan
Keywords:
Human-agent Interaction, Human Factors, Decision-making, Work Engagement.
Abstract:
When human participants collaborate in decision-making and problem solving, the results are often better than
those obtained from individual efforts. However, when they cooperatively perform tasks with an embodied
agent, the agent is often regarded merely as a human-centric multi-modal interface from which information is
obtained. In this study, by increasing the “personal resource” in the aspect of work engagement, we aim to
incorporate active attitude in the human participants toward task and in their interactions with the embodied
agents. We conducted an experiment to investigate whether increasing the personal resource will impact the
active attitude of participants. In the experiment, we used a mediator agent that increased either the “personal
resource” or “job resource” in addition to an expert agent that directly supported the task. The results suggested
that a mediator agent that increased the personal resources can induce the active attitude of participants in the
human-agent interaction.
1 INTRODUCTION
Agents capable of intelligent response are expected
not only to act as multimodal interfaces that provide
solutions to problems efficiently but also to facilitate
more creative activities by promoting human under-
standing and learning. When the problems are collab-
oratively solved, we expect that the results are often
better than those obtained from individual efforts.
In ICAP theory, in the framework of human co-
operative learning (Chi and Wylie, 2014), the activity
states of learners in cooperative learning are catego-
rized under the following four states: 1) Passive state:
passively learning from a teacher; 2) Active state: ac-
tively and voluntarily tackling the problem; 3) Con-
structive state: reconstruction of metacognition and
own knowledge through discussion with others; 4) In-
teractive state: reconstruction of knowledge through
criticism and refutation of others’ ideas. It was dis-
covered that the performance of the learners in coop-
erative learning increased as they transitioned to the
interactive state (I > C > A > P). When an agent
efficiently searches for a solution, and returns a re-
sponse to a human question, the human is in the pas-
sive state. When collaborating with agents on tasks,
the goal should be to attain the interactive state in the
ICAP theory. However, many cooperative agents have
not been able to effectively change the human state
from Passive to Active (Raux et al., 2005; Misu et al.,
2011). In order to induce humans to actively and vol-
untarily tackle problems through human agent inter-
action, it is necessary to elicit an active response from
humans to the agent’s interactions.
The main components of an agent that differ
from those of other systems are summarized by re-
searchers, such as (Wooldridge, 2009; Russell and
Norvig, 2002). From those, we consider that it is im-
portant for agents to demonstrate ”reactivity” to hu-
mans and ”social ability”. ”Reactivity” is the ability
to incorporate information from others into one’s own
behavior. ”Social ability” is the ability to assume that
others have intentions of sharing information and out-
comes. Meanwhile, to recognize these elements, and
induce an ”intentional stance, it is important a hu-
man’s active attitude which tries to deduce the inner
state of an agent.
As we expect humans and agents to collaborate
to accomplish tasks, we focused on “work engage-
ment” as a form of active attitude in task execution.
Schaufeli et al. (Schaufeli and Bakker, 2004) de-
scribed high work engagement as a state of mind
associated with a positive, fulfilling sense of work.
Employees with a high level of work engagement
are enthusiastic about their work, and tend to regard
Ohmoto, Y., Kuno, M. and Nishida, T.
Improving Active Attitude for Interactive Decision-making with Multiple Agents by Increasing Personal Resource.
DOI: 10.5220/0010265503290336
In Proceedings of the 13th International Conference on Agents and Artificial Intelligence (ICAART 2021) - Volume 1, pages 329-336
ISBN: 978-989-758-484-8
Copyright
c
2021 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
329
work as rewarding rather than stressful. Bakker et
al. (Bakker et al., 2004) demonstrated that employ-
ees with high work engagement were highly rated by
their colleagues for their performances in both obli-
gations and volunteer work.
Kotze (Kotze, 2018) highlighted two main types
of work engagement resources as follows. 1.) Job
resources: The characteristics of a job that facilitate
the realization of a job’s goals, and drive the growth
of those involved. 2.) Personal resources: A positive
self-esteem associated with an individual’s sense of
control and influence over his/her environment. Job
resource can be regarded as the resource that exter-
nally motivates a person, and personal resource as one
that internally motivates a person. Increasing the re-
sources enhances work engagement. We think that
job resources are related to “reactivity, whereas per-
sonal resources are related to “social ability.
The framework that focuses on work engagement
is based on studies concerning workers in a day-to-
day work environment. We considered the conver-
sation and interaction were regarded as basic activi-
ties (i.e., day-to-day work) for performing the task.
Therefore, in this study, we assume that by enhancing
work engagement in decision-making tasks, it is pos-
sible to increase the active tackling of a human toward
the task and interaction with an agent.
The purpose of this study is to investigate whether
the application of a work engagement enhancement
framework to a cooperative decision-making task
with agents can make human responses to agents
more active. Our final purpose is to improve the at-
titude of people working on cooperative tasks with
agents such that they attain the interactive state in
the ICAP theory. The goal of this research is to in-
duce Active” human attitudes in human-agent inter-
actions, which are the most fundamental elements of
cooperative decision-making, as a first step in achiev-
ing this goal. A human attitude in performing tasks
can be either explicitly observed from one’s behav-
ior or implicitly embedded in it. Physiological in-
dices are measured to observe changes in human inner
states that are not easily inferred from the behavior
during human-agent interactions; subsequently, these
are used in addition to the explicitly expressed behav-
ior for evaluation.
The present paper is organized as follows. In next
section, we briefly explains the multi-agent system
designed with the method to increase resources re-
lated to work engagement. In the Experiment sec-
tion, we presents the results of an experiment con-
ducted to investigate the effect of a specific behavior
of the agent that increased the personal resources. In
the Discussion section, the contributions of this study
and a future study are described. The conclusions are
presented in the Conclusion section.
2 CONSULTATION AGENTS TO
INCREASE JOB AND
PERSONAL RESOURCES
For humans to perform decision-making interactively
with an agent, some methods were developed to
induce the “intentional stance”(Dennett, 1988) by
presenting goal-oriented behaviors, expressing con-
tingent responses, and dynamically estimating the
user’s behavioral intentions in previous studies (e.g.,
(Ohmoto et al., 2017)). In these methods, the agent
enables the human to recognize its “social ability” by
demonstrating “reactivity”. Meanwhile, to maintain
the “intentional stance,” it is still important that a hu-
man attempts to actively estimate the internal state of
the agent.
In the tasks that humans and agents collabora-
tively accomplish, we focused on “work engagement”
as a form of active attitude in task execution. In con-
trast to previous studies, we investigate whether work
engagement, a type of “reactivity” to the agent’s ac-
tivities, is induced by the agent’s manifestation of “so-
cial ability. Specifically, between “job resource” and
“personal resource” that have been proposed as re-
sources to enhance work engagement, we investigate
whether agents can improve humans’ active attitude
by revealing the agent behavior designed to increase
the “personal resource” related to social relationships.
2.1 Agent Behavior Related to
Increasing Resources
Job resources typically include interactions that facil-
itate task accomplishment and ability improvement,
such as peer support and performance feedback. Per-
sonal resources typically include interactions involv-
ing social relationships with others, such as enhancing
self-efficacy and self-esteem.
The basic function of consultation agents is to pro-
pose and explain solutions, methods, and hints. This
implies that the consultation agent functions as “a
colleague who helps in decision-making” to increase
job resources. In other words, existing consultation
agents increase the job resources to stimulate human
interactions and tasks.
Some agents are designed to increase personal re-
sources. For example, studies that attempt to gener-
ate rapport between a human and agent (e.g., (Gratch
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
330
et al., 2007)) have been conducted. These agents at-
tempt to encourage human interactions and tasks by
improving social relationships.
To make a decision in an unfamiliar domain, an
expert’s opinion is needed. The expert contributes to
increased job resource. On the other hand, the behav-
ior to increase personal resource is sometimes conflict
the behavior to increase job resource. Therefore, con-
ducting an interaction with the goal of increasing per-
sonal resources using only one agent is challenging.
In this study, we thus used two types of agents: (i) an
expert agent with advanced knowledge of tasks; (ii) a
mediator agent who has only a general knowledge of
tasks. We also expect the participant in the interaction
to estimate the agent interaction model by observing
the interaction between the agents. This is similar to
the communication strategy used to persuade users ef-
fectively (Walster and Festinger, 1962; Andr
´
e et al.,
2000). Although we do not focus on persuasion in the
current study, it is useful for consultation interaction.
2.1.1 Role of an Expert Agent
The role of the expert agent is to answer user ques-
tions, and generate a proposal that will aid the user
(and a mediator agent) in decision-making. The ac-
tions performed by an expert agent are as follows:
Provide responses for the user by employing filler
and nodding motions.
Provide responses for the mediator agent based on
the predefined responses.
Provide proposals, which are similar to the user
preferences, from a prepared list.
Modify the contents of the proposals, and provide
variations that are similar to the user preferences,
from the prepared list.
Explain three major factors included in the pro-
posals.
Answer the user questions.
2.1.2 Role of a Mediator Agent
The role of the mediator agent is to address and com-
ment on some of the questions and requests that a user
with only a general knowledge of the tasks may have.
In this study, the mediator agent additionally performs
an action to increase either the personal resources or
job resources. The actions typically performed by the
mediator agent are as follows:
Provide responses for the user by employing filler
and nodding motions.
Respond to the expert agent based on the prede-
fined responses.
Provide questions that are similar to the user pref-
erences to the expert agent, from the prepared list.
Provide comments related to decision-making
factors that have not been considered by the user
and expert agent.
The mediator agent has knowledge of the major as-
pects of the decision-making. The questions and com-
ments by the mediator agent are focused on a few fac-
tors included in the proposals for decision-making.
When the mediator agent is configured to take ac-
tions that increase the user’s personal resources, it
performs the following actions additionally:
Agree with the user’s opinions and requests.
Concur on the aspects with which the user is pre-
occupied.
Compliment the user’s decision making process.
Nod elaborately after listening to the user’s opin-
ions and requests.
All these actions performed to increase personal re-
sources are independent of the proposal content. We
set the probability of action such that these actions do
not occur in succession and not appear contrived.
When the mediator agent is configured to perform
actions that increase the user’s job resources, it per-
forms the following actions additionally:
Provide meta-cognitive suggestions regarding the
user opinions and requests, such as “There are
other forms of thinking.
Provide comments regarding aspects that are dis-
regarded by the user.
Advise the user on decision making.
Display thoughtful gestures after listening to the
user’s opinions and requests.
Actions that increase job resources are determined by
the emphasizing factors estimated from the user be-
havior. The advisory-like behaviors occur when the
emphasis factor can be estimated with a certain level
of confidence.
2.2 Outline of the Interactive
Decision-making with the Agents
Figure 1 shows the components of the agents and the
outline of the data flow. Both the mediator agent and
the expert agent estimate the user preferences, and
provide proposals and behaviors using the dynamic
estimation of emphasizing points (DEEP) proposed
in previous studies. DEEP is a method for estimating
the user’s emphasis factors based on verbal reactions,
body movements, and physiological indices (Ohmoto
et al., 2014). The physiological indices used in this
study are skin conductance response (SCR) and heart
Improving Active Attitude for Interactive Decision-making with Multiple Agents by Increasing Personal Resource
331
User
A mediator agent An expert agent
Estimation of
emphasis points
Estimation of
emphasis points
Nonverbal
Behavior
Decision
Utterance
Decision
Nonverbal
Behavior
Decision
Utterance
Decision
Proposal
Decision
Typical reactions
for the interaction
Typical reactions
for the interaction
Explanations of the
proposal
Actions for increasing
the job/personal
resources
physiological indices
physiological indices
utterance and
nonverbal behavior
utterance and
nonverbal behavior
Figure 1: The components of the agents and the outline of
the data flow.
rate (this is converted to cardiac sympathetic index
(CSI) and cardiac vagal index (CVI)). The SCR be-
comes unresponsive when a person concentrates on
one object, and often becomes responsive when a per-
son focuses on various objects. The CSI reflects the
state of the sympathetic nervous system, and partic-
ipants with a high CSI are relatively tensed and ex-
cited. Furthermore, the CVI reflects the state of the
parasympathetic nervous system, suggesting that par-
ticipants with a higher CVI tend to be less tensed and
more relaxed. Using these physiological indices as
clue, DEEP estimates the user’s emphasis factors in
addition to the verbal and nonverbal information.
The proposals by the expert agent are categorized
into approximately ten categories, based on the major
factors for decision-making included in the propos-
als. The user repeatedly asks questions regarding the
proposal, and continues to evaluate the proposals pro-
vided by the agent until satisfaction is attained. The
expert agent’s potential proposal contents are selected
using a pre-prepared table. A few proposals that are
close to the combination of factors are prepared for
when the user requests a variation of the proposal with
similar factors. The answers to the user’s questions
are prepared in advance based on preliminary interac-
tions. If an appropriate answer is not included in the
list, the expert agent apologizes, and asks whether the
user has another question.
In addition, when some proposal candidates ap-
pear, the selecting method is different between for in-
creasing the personal resource and for increasing the
job resource. When the agent tries to increase the per-
sonal resource, the proposal is made to conform to the
user’s most recent preference. When the agent tries to
increase the job resource, the proposal is tailored to
give the user a relatively wide viewpoint.
We used Unity3D (http://www.unity.com/) as the
interface for the interaction agents. The agent could
automatically realize multimodal interaction behav-
iors, such as gaze, body motions, and speech. The
data related to verbal meanings were manually pro-
vided to the agents, such as whether a user response
was positive or negative and what question was asked
by the user, as these could not be efficiently deter-
mined in real time. Although the data related to ver-
bal meanings were manually provided, the agent au-
tomatically generated the output of verbal and non-
verbal behaviors that had been previously designed,
except for the answers to unexpected questions. We
refer to the agent control method using partially man-
ual inputs as the Wizard of Oz (WoZ).
The WoZ operator uses two commands: decision
of speech contents and change in proposal presenta-
tion. When the user makes an utterance and expects
a meaningful response from the agents, the operator
determines which agent should respond, and selects a
corresponding response from a prepared response list.
Subsequently, a command is returned to the agent.
The agent who receives the command provides the
selected response to the user. When the user requests
to change the proposal presented, the operator sends
the corresponding command to change the proposal.
The agent automatically provides some prepared re-
sponses for changing the proposal, and then changes
the proposal.
3 EXPERIMENT
The purpose of this experiment is to investigate the
effect of an agent behavior, which increased personal
resources for improving the active attitude of the par-
ticipants. In the experiment, the participants were
asked to plan a two-day trip to a fictitious city with
two agents (an expert agent and a mediator agent).
The participants were randomly divided into three
groups based on the behavior of the mediator agent.
In the personal resource group (PR-group), the me-
diator agent’s function was to increase personal re-
sources. In the job resource group (JR-group), the
mediator agent’s function was to increase the job re-
sources. No mediator agent was involved in the no-
agent group (NA-group). The behavior of the expert
agent toward the participants was the same across the
groups. By comparing the behavioral data of these
groups, we investigate the effects of the agent’s inter-
actions on the participants’ active attitude.
3.1 Task
Planning the two-day trip involved the following: de-
termining three tourist destinations to visit, and the
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
332
Agent control PC
WoZ
environment
Display the agents’ behavior
Experimenter
Participant
Commands
Figure 2: The experimental setting.
route and order in which to visit them; arranging ac-
commodations and the length of stay at accommoda-
tion. Each destination was divided into six categories
depending on the factors that contributed to one’s en-
joyment, such as “sweets” and “sightseeing.
The travel planning tasks were divided into three
phases: introduction phase, overview phase, and de-
tail phase. In the introduction phase, the participant
conversed with the agents. The conversation included
a self-introduction and explanation on how to proceed
with the task. In the overview phase, the participant
considered various sightseeing spots, tourist destina-
tions, and inns to determine the three destinations that
the participant wanted to visit. The overview phase
ended when the participant was satisfied. In the detail
phase, the route to travel was first determined by con-
sidering the location of the tourist destinations and the
accommodation decided in the overview phase. Sub-
sequently, the participant could customize the details
of his/her travel plan based on the route. The detail
phase ended when the participant was satisfied with
the customization.
In this task, the mediator agent’s behavior was dif-
ferent for each group. In the PR-group, the content of
the destination proposed by the mediator agent em-
phasized the nearest preference of the participant. In
the JR-group, the mediator agent increased the job re-
sources for the participant. The content of the destina-
tion proposed by the mediator agent included factors
that were relevant to the preferences of the participant
but had not been considered so far. In the NA-group,
no mediator agent was involved (a participant and an
expert agent collaborated on the task).
3.2 Experimental Setting
The experimental setting is shown in Figure 2. The
participant was seated in front of a 60-inch monitor
displaying the agents and the map of the travel des-
tinations. The participant’s voice was recorded us-
ing microphones, and the participant’s behavior was
recorded using two video cameras. The participants
interacted with the agents using only his/her voice. To
estimate the mental state of the participant, the SCR
and heart rate were recorded using a device (Polymate
mini). These were sent to the agents in real time.
After a brief explanation of the experiment, elec-
trodes for measuring the SCR and heart rate were at-
tached to the participant’s left hand and chest. Af-
ter a 2-min relaxation period, the experimenter turned
on the video cameras and commenced recording of
the physiological indices. The participants performed
the three phases of human-agent interactions sequen-
tially. The total time of the experiment was approxi-
mately 1 h.
The participants in the experiments were 45
Japanese undergraduate college students (27 men and
18 women). The average age was 23.2 years (standard
deviation 3.28). The participants whose physiological
indices could not be accurately obtained were elimi-
nated. The PR-, JR-, and NA-groups included 13, 15,
and 14 participants, respectively.
3.3 Result
3.3.1 Reaction Latency of Utterances
To induce a user/person’s active attitude towards the
interaction with agents, we focused on an interaction
cue which is related to the mental states of people ob-
servable in conversations. Bechade et al. explored
behavioral spoken cues in human-robot interaction
(Bechade et al., 2015). In the study of Bechade et
al., they found a significant interaction between self-
confidence and the number of speech segments, self-
confidence and the speech reaction time, and enthu-
siasm and the participant speech duration. Ono et al.
focused on synchronous behaviors of two robots and
a person close to two robots became easily involved in
their communication (Ono et al., 2016). In the study
of Ono et al., they focused on overlapping interaction.
The study also suggested that participants felt that the
degree of communication activity became lively by
presenting the overlapping interaction.
From the cues considered in these studies, we
analysed reaction latency of utterances for confirming
the active attitudes of the participant. We expected
that a participant of interaction with an agent which
showed the short reaction latency would feel active
engagement of the agent towards the conversation.
The reaction latency from the end of the agent’s
utterance to the start of the participant’s utterance was
measured. The filler was not regarded as an utter-
ance. Figure 3 shows the average of the latency in
the task. We performed a 2 (phase: overview or de-
tail) x 3 (group: PR, JR, or NA) analysis of variance
Improving Active Attitude for Interactive Decision-making with Multiple Agents by Increasing Personal Resource
333
1
2
3
4
5
overview detail
PR JR NA
The latency of
utterances
in the task
Figure 3: The latency of utterances in the task.
5
6
7
8
9
overview detail
PR JR NA
The
number
of changes
of
emphasizing
factors
Figure 4: The number of changes of emphasizing factors.
(ANOVA) on the data of the reaction latency of the
participants’ utterance. Consequently, significant dif-
ferences were observed among the phases and groups
(phase: F(1, 39) = 36.39, p < 0.0001; group: F(2, 39)
= 4.99, p = 0.012).
We performed multiple comparisons among the
groups using Ryan’s method. Significant differences
were shown between the PR- and JR-groups (PR <
JR, p = 0.0050), and between the PR- and NA-groups
(PR < NA, p = 0.023). Although the reaction la-
tency became longer in the detail phase across all
the groups, the PR-group remained at the level of the
overview phase in the other groups. Therefore, we
infer that by increasing the personal resources, the
reaction latency of the participant can remain short,
without being affected by the decision-making con-
tent. The results suggest that we were able to in-
duce active attitudes towards the interaction with the
agents to some extent. In addition, it is important to
predict the behavior of others for remaining the reac-
tion latency of utterance short. We perhaps infer that
increasing the participant’s personal resources facili-
tates the estimation of the agent’s behavioral model.
3.3.2 Emphasizing Factors in the
Decision-making
In this task, DEEP was used to perform an interac-
tion by dynamically estimating the participant’s em-
phasis factors. To investigate the effect of interacting
with the agents on the decision-making, we counted
the number of times that a participant’s prioritization
of the emphasizing factors changed during the task.
Figure 4 shows the average of the number of changes.
We performed a 2 (phase: overview or detail) x 3
(group: PR, JR, or NA) ANOVA on the data of the
2
3
4
5
overview detail
PR JR NA
The
number of
the detailed considerations
Figure 5: The number of the detailed considerations.
number of times that a participant’s prioritization of
the emphasis factors changed. Consequently, a sig-
nificant difference in phase was observed (F (1, 39) =
8.83; p = 0.0051). The paired t-tests in each group be-
tween the overview and detail phases revealed signif-
icant reductions in the JR- and NA-groups (JR: t (14)
= 3.09, p = 0.008; NA: t (13) = 2.7; p = 0.018) but not
in the PR-group (t (12) = 0; p = 1.0). This implies
that the participants in the JR- and NA-groups did
not change their preferences in the detail phase. The
participants in the PR-group did not adhere to their
own preferences even in the detail phase; they contin-
uously considered their emphasis factors throughout
the interaction with the agents. This suggests that the
increased personal resources of the participants en-
sured that they maintained their active attitude toward
improving their travel plans in the detail phase.
To confirm this, we counted the number of times
that the participants considered parts that were not
mentioned by the agent. Figure 5 shows the average
number of the detailed considerations. We performed
a 2 (phase: overview or detail) x 3 (group: PR, JR, or
NA) ANOVA on the data of the number of times of the
detailed considerations. Consequently, a significant
difference among the phases was observed (F(1, 39)
= 7.28; p = 0.0010). The paired t-tests in each group
between the overview and detail phases revealed no
significant increases in the specified values in the JR-
and NA-groups (JR: t (14) = -1.13, p = 0.28; NA: t(13)
= -1.20, p = 0.25) but a significant increase in the PR-
group (t(12) = -2.23, p = 0.046). Therefore, we in-
ferred that increasing the personal resources induced
the active attitude of the participants toward the task.
3.3.3 Physiological Indices
To investigate whether a change that was not appar-
ent from the behavior occurred in the human men-
tal state, we analyzed the heart rate indices (CSI and
CVI). CSI is one of the indices of sympathetic nerve
activity. The sympathetic nervous system’s primary
function is to stimulate the body’s fight-or-flight re-
sponse, such as tension and excitement. The CVI is
one of the indices of parasympathetic nerve activity.
The parasympathetic system is responsible for stimu-
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
334
1.5
1.6
1.7
1.8
1.9
2
overview detail
PR JR NA
The
average
of CSI values
Figure 6: The number of times of reversed between SCR to
CSI and CSI to SCR.
The
average
of CVI values
4.5
5
5.5
6
6.5
overview detail
PR JR NA
Figure 7: The number of times of reversed between SCR to
CVI and CVI to SCR.
lating the “rest-and-digest” activities that occur when
the body is at rest and relaxed.
We calculated the CSI values in each phase
(overview or detail) for each group. The results are
shown in Fig. 6. We performed a 2 (phase: overview
or detail) x 3 (group: PR, JR, or NA) ANOVA on the
data of the CSI values. Consequently, a significant
difference was observed across the phases (F (1, 39)
= 5.42, p = 0.025). The paired t-tests results in each
group in both the overview and detail phases show no
significant increases in the PR- and NA-groups (PR: t
(12) = -0.52; p = 0.61; NA: t(13) = -1.55; p = 0.14);
however, a significant increase was observed in the
JR-group (t (13) = -2.33; p = 0.035). These results
suggest that the tension and excitement of the partici-
pants increased in the detail phase in the JR-group.
We also calculated CVI values in each phase
(overview or detail) in each group. The results are
shown in Fig. 7. We performed a 2 (phase: overview
or detail) x 3 (group: PR, JR, or NA) ANOVA on the
data of the CVI values. There was no significant dif-
ference across the groups and phases. This implies
that the actions of the mediator agent directed toward
increasing the job or personal resources did not affect
the participants mentally in both the rest and relaxed
states.
4 DISCUSSION
In the experiment, the mediator agent increased the
personal resources and job resources in the PR- and
JR-groups, respectively. The expert agent increased
the job resources in any group. According to Xan-
thopoulou et al. (Xanthopoulou et al., 2009), the two
resources are interrelated, and when both resources
are fully acquired, they influence each other, and in-
duce higher work engagement. Therefore, in this
study, it is assumed that both the expert agent that in-
creased the job resources and the mediator agent that
increased the personal resources are important factors
toward increasing a participant’s active attitude in in-
teracting with the agents. As the results of the analy-
sis, it is considered that, by the agent which increases
the personal resource, the active attitude of the par-
ticipants was induced to some extent both for the in-
teraction with the agents and for the execution of the
task.
However, no difference was found in the active at-
titude of the participant based on whether there is a
mediator agent increasing the job resources. This im-
plies that as the job resource is increased to a certain
limit, and further actions did not significantly influ-
ence the participant’s attitude. On the other hand,
there was a significant difference in the value of the
physiological index (CSI), and the mediating agent
that increased the job resource tended to enhance
a participant’s stress. Considering the participants’
stress levels, the effect of the support by the agents
is not always positive.
We performed analysis to identify the changes
in the human participants’ inner states using physi-
ological indices (CSI and CVI in the result section).
However, there were diverse stimuli in the experi-
ment, and the human inner state varies were contin-
uously affected by these stimuli. To eliminate these
effects, we focused on the causal relationship between
the changes in several physiological indices. The
“CausalImpact” package of R 3.6.0 was used to es-
timate the causality. The analysis in the result sec-
tion shows that the CSI is affected by the agent’s ac-
tions directed toward increasing the job or personal
resource. Although the result was not demonstrated,
no significant difference was observed in the SCR
measured in addition to the CSI. We thus examined
the effect of the agent influence on the change in
the CSI using the SCR value as a covariate. The
CausalImpact package was used to calculate the value
of the point-wise impact in the 30-s data before and
after the agent’s action directed toward increasing the
job or personal resource. A t-test was performed to
determine whether the values changed following the
agent’s action differed between the PR-group and JR-
group. There was a significant difference (t (161) =
-2.05, p = 0.042; JR > PR). From the analysis it can
be deduced that the intensity of the effect of the agent
actions on the CSI can be observed, taking into ac-
count the change in the participant’s mental state, as
Improving Active Attitude for Interactive Decision-making with Multiple Agents by Increasing Personal Resource
335
reflected in the SCR. By estimating the causal rela-
tionships between the variables that serve as clues to
inferring the changes in a person’s mental state, we
may be able to compare the inferred changes while
eliminating various convoluted factors in future work.
5 CONCLUSION
The aim of this study was to investigate whether an
active attitude can be induced in a person by apply-
ing the framework of work engagement enhancement
to a cooperative decision-making task with agents.
We conducted experiments to evaluate the effect of
an agent behavior directed toward increasing personal
resources for improving the active attitude of the par-
ticipants. In the experiment, the participants were
asked to plan a two-day trip to a fictitious city with
two agents (an expert agent and a mediator agent).
There were three groups; PR-group: the mediator
agent serves to increase personal resources, JR-group:
the mediator agent serves to increase job resources,
and NA-group: no mediator agent is involved. From
the results, we suggest that the mediator agent can in-
duce and maintain the participant’s active attitude to-
ward the task and the agents by encouraging the par-
ticipant to increase his/her personal resources. In fu-
ture work, we will consider a method for analyzing
the detailed inner state changes of the participants in-
volved in the task.
ACKNOWLEDGEMENTS
This research is supported by Grant-in-Aid for Young
Scientists (B) (KAKENHI No. 16K21113), and
Grant-in-Aid for Scientific Research on Innovative
Areas (KAKENHI No. 26118002) from the Ministry
of Education, Culture, Sports, Science and Technol-
ogy of Japan.
REFERENCES
Andr
´
e, E., Rist, T., Van Mulken, S., Klesen, M., and Baldes,
S. (2000). The automated design of believable di-
alogues for animated presentation teams. Embodied
conversational agents, pages 220–255.
Bakker, A. B., Demerouti, E., and Verbeke, W. (2004).
Using the job demands-resources model to predict
burnout and performance. Human Resource Manage-
ment: Published in Cooperation with the School of
Business Administration, The University of Michigan
and in alliance with the Society of Human Resources
Management, 43(1):83–104.
Bechade, L., Dubuisson Duplessis, G., Sehili, M., and Dev-
illers, L. (2015). Behavioral and emotional spoken
cues related to mental states in human-robot social in-
teraction. In Proceedings of the 2015 ACM on Interna-
tional Conference on Multimodal Interaction, pages
347–350. ACM.
Chi, M. T. and Wylie, R. (2014). The icap framework: Link-
ing cognitive engagement to active learning outcomes.
Educational psychologist, 49(4):219–243.
Dennett, D. C. (1988). The intentional stance in theory and
practice.
Gratch, J., Wang, N., Gerten, J., Fast, E., and Duffy, R.
(2007). Creating rapport with virtual agents. In In-
ternational Workshop on Intelligent Virtual Agents,
pages 125–138. Springer.
Kotze, M. (2018). How job resources and personal
resources influence work engagement and burnout.
African Journal of Economic and Management Stud-
ies, 9(2):148–164.
Misu, T., Sugiura, K., Kawahara, T., Ohtake, K., Hori, C.,
Kashioka, H., Kawai, H., and Nakamura, S. (2011).
Modeling spoken decision support dialogue and opti-
mization of its dialogue strategy. ACM Transactions
on Speech and Language Processing (TSLP), 7(3):10.
Ohmoto, Y., Kataoka, M., and Nishida, T. (2014). The ef-
fect of convergent interaction using subjective opin-
ions in the decision-making process. In Proceedings
of the Annual Meeting of the Cognitive Science Soci-
ety, volume 36.
Ohmoto, Y., Ueno, S., and Nishida, T. (2017). Effect of an
agent’s contingent responses on maintaining an inten-
tional stance. In Proceedings of the 5th International
Conference on Human Agent Interaction, pages 277–
285. ACM.
Ono, T., Ichijo, T., and Munekata, N. (2016). Emergence
of joint attention between two robots and human us-
ing communication activity caused by synchronous
behaviors. In Robot and Human Interactive Com-
munication (RO-MAN), 2016 25th IEEE International
Symposium on, pages 1187–1190. IEEE.
Raux, A., Langner, B., Bohus, D., Black, A. W., and Es-
kenazi, M. (2005). Let’s go public! taking a spoken
dialog system to the real world. In Ninth European
Conference on Speech Communication and Technol-
ogy.
Russell, S. and Norvig, P. (2002). Artificial intelligence: a
modern approach.
Schaufeli, W. B. and Bakker, A. B. (2004). Job demands,
job resources, and their relationship with burnout and
engagement: A multi-sample study. Journal of Or-
ganizational Behavior: The International Journal of
Industrial, Occupational and Organizational Psychol-
ogy and Behavior, 25(3):293–315.
Walster, E. and Festinger, L. (1962). The effectiveness of”
overheard” persuasive communications. The Journal
of Abnormal and Social Psychology, 65(6):395.
Wooldridge, M. (2009). An introduction to multiagent sys-
tems. John Wiley & Sons.
Xanthopoulou, D., Bakker, A. B., Demerouti, E., and
Schaufeli, W. B. (2009). Reciprocal relation-
ships between job resources, personal resources, and
work engagement. Journal of Vocational behavior,
74(3):235–244.
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
336