Multi-Agent System for Emulating Personality Traits Using Deep Reinforcement Learning

Abstract

Conventional personality assessment methods depend on subjective input, while game-based AI predictive methods offer a dynamic and objective framework. However, training these models requires large and labeled datasets, which are challenging to obtain from real players with diverse personality traits. In this paper, we propose a multi-agent system using Deep Reinforcement Learning in a game environment to generate the necessary labeled data. Each agent is trained with custom reward functions based on the HiDAC system that encourages trait-aligned behaviors to emulate specific personality traits based on the OCEAN personality trait model. The Multi-Agent Posthumous Credit Assignment (MA-POCA) algorithm facilitates continuous learning, allowing agents to emulate behaviors through self-play. The resulting gameplay data provide diverse, high-quality samples. This approach allows for robust individual and team assessments, as agent interactions reveal the impact of personality traits on team dynamics and performance. Ultimately, this methodology provides a scalable, unbiased methodology for human personality evaluation in various settings, establishing new standards for data-driven assessment methods.

Full text

Citation: Liapis, G.; Vlahavas, I.;
Multi-Agent System for Emulating
Personality Traits Using Deep
Reinforcement Learning. Appl. Sci.
2024,14, 12068. https://doi.org/
10.3390/app142412068
Academic Editors: Rakib Abdur and
Mehmet Aydin
Received: 15 November 2024
Revised: 19 December 2024
Accepted: 22 December 2024
Published: 23 December 2024
Copyright: © 2024 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Article
Multi-Agent System for Emulating Personality Traits Using Deep
Reinforcement Learning †
Georgios Liapis ‡and Ioannis Vlahavas *,‡
School of Informatics, Aristotle University of Thessaloniki, 541 24 Thessaloniki, Greece; gliapisa@csd.auth.gr
*Correspondence: vlahavas@csd.auth.gr; Tel.: +30-694-447-0170
†
This paper is a revised and expanded version of a paper entitled: Liapis, G.; Vordou, A.V.I. Machine Learning
Methods for Emulating Personality Traits in a Gamified Environment. In Proceedings of the 13th Conference
on Artificial Intelligence (SETN 2024), Piraeus, Greece, 11–13 September 2024.
‡These authors contributed equally to this work.
Abstract: Conventional personality assessment methods depend on subjective input, while game-
based AI predictive methods offer a dynamic and objective framework. However, training these
models requires large and labeled datasets, which are challenging to obtain from real players with
diverse personality traits. In this paper, we propose a multi-agent system using Deep Reinforcement
Learning in a game environment to generate the necessary labeled data. Each agent is trained with cus-
tom reward functions based on the HiDAC system that encourages trait-aligned behaviors to emulate
specific personality traits based on the OCEAN personality trait model. The Multi-Agent Posthumous
Credit Assignment (MA-POCA) algorithm facilitates continuous learning, allowing agents to emulate
behaviors through self-play. The resulting gameplay data provide diverse, high-quality samples. This
approach allows for robust individual and team assessments, as agent interactions reveal the impact
of personality traits on team dynamics and performance. Ultimately, this methodology provides a
scalable, unbiased methodology for human personality evaluation in various settings, establishing
new standards for data-driven assessment methods.
Keywords: machine learning; serious games; personality assessment; multi-agent
1. Introduction
Personality, which can be defined under characteristic patterns of thought, emotion,
and behavior that remain consistent over time and situations [
1
], plays a key role in
shaping behavior, thoughts, and emotions, making it valuable to understand in personal
and professional contexts for insights into actions, decisions, and growth. Self-report
questionnaires and psychological assessments are common tools for evaluating personality,
but they come with limitations, as self-perception biases may affect accuracy. These tools
are best complemented by additional feedback or professional evaluations, which can be
enhanced by technology [2].
Advancements, especially in gaming, have expanded into fields like education, as-
sessment, and diagnosis. This has led to “serious games”, which use gaming for practical
purposes like skill-building and evaluation. Gaming environments promote exploration,
problem-solving, and skill development, enhancing cognitive and soft skills like critical
thinking and adaptability, which are valuable in today’s workplace [3].
Escape Room (ER) games, a genre that emphasizes teamwork and communication,
have emerged as tools in corporate settings for evaluating and fostering these skills. ER
games combine physical and mental challenges in dynamic settings, allowing personality
traits to manifest naturally and offering a richer understanding than static assessments.
However, traditional evaluations of team performance in ER games rely on post-game
questionnaires, which can be biased and fail to capture all player actions [
4
]. A digital
Appl. Sci. 2024,14, 12068. https://doi.org/10.3390/app142412068 https://www.mdpi.com/journal/applsci

Appl. Sci. 2024,14, 12068 2 of 20
Escape Room game could overcome these limitations, providing a more comprehensive,
real-time approach to tracking and evaluating individual and team dynamics [5].
So, emulating human behavior in a gaming environment is a complex undertaking
due to the complexity of personality, which makes it challenging to quantify without a
large number of distinct human players with specific traits.
To develop a new personality assessment method, we can combine a digital, serious
Escape Room (ER) game with AI technology. This setup allows the AI to analyze gameplay
data, identifying patterns that improve assessment accuracy. The ER game triggers human
player behaviors, and an AI-based regression system analyzes the data collected to assess
their personality traits following the OCEAN Five model [
6
], which was chosen due
to being the most scientifically robust and widely used model in academic psychology
and research. However, training this Multi-Output Regression System (MORS) requires
extensive labeled data, which is challenging to obtain from human participants with specific
trait assessments.
To overcome the challenge of gathering large amounts of labeled personality data, we
developed a multi-agent system that can simulate gameplay and generate the necessary
data through self-play. Each agent is trained on a specific personality trait, based on custom
rewards functions that are based on mathematical formulas from the HiDAC model [
7
]
that represents behaviors. The agents are rewarded not only for solving challenges within
the environment but also for displaying behaviors that align with their assigned traits
while cooperating.
Related simulations and their results can serve as a base and a benchmark for modeling
behaviors in other environments. In this study, we used the HiDAC crowd simulation
as our baseline, which is frequently utilized as a benchmark for evaluating models of
human behavior.
To train the agents effectively, we used the MA-POCA (Multi-Agent Posthumous
Credit Assignment) [
8
] reinforcement learning algorithm. This approach enables agents
to continue learning even after they complete the environment—in this case, when they
“escape” the game. This simulation-based training with MA-POCA provides an efficient
way to gather high-quality data, refining the MORS’s ability to recognize subtle human
behavioral patterns and enhancing the overall accuracy of personality assessments within
the Escape Room game environment. By assigning agents a range of personality traits,
from high to low, and rewarding trait-aligned behaviors, the system generates a diverse set
of gameplay data that improve the MORS’s ability to assess personality accurately.
The data generated through this process are then used to train the MORS utilizing
supervised algorithms. Consequently, when human players interact with the game, their
gameplay data are gathered and fed to the MORS so that it can accurately assess their
personality based on the behaviors exhibited during the game, as shown in Figure 1.
Though in this paper, we will only delve into the multi-agent implementation, where we
showcase how the final system works.
The emulation of behaviors not only for a single agent [
9
] or team-related traits [
10
]
but also for all possible traits and behaviors is crucial for generating valuable and high-
quality data. Moreover, our approach allows us to evaluate the effectiveness of a team of
agents based on each agent’s personality attributes and gameplay style. This holistic view
enables a deeper understanding of how different personality traits interact and contribute
to overall team dynamics and performance.

Appl. Sci. 2024,14, 12068 3 of 20
Figure 1. The workflow of the MindEscape environment (the colored entities are the ones analyzed
in this paper).
This paper presents a multi-agent system in a 3D digital ER environment as a simula-
tion of individual and team effectiveness based on the behaviors and personalities of each
team member. We propose a specific methodology for the reward functions so that each
agent acts and emulates behaviors based on the specific OCEAN Five personality traits
model. This methodology can also be used in other types of games and scenarios.
The proposed contributions are as follows:
•
A reward methodology for emulating human behaviors in a dynamic gamified envi-
ronment;
•
A multi-agent system that measures team efficiency based on the personality traits of
the team member;
•
A way to generate synthetic behavioral data using deep reinforcement learning agents
through self-play.
2. Materials and Methods
In this section, we analyze the core game mechanics, how the escape room environ-
ment works, the main components of the agents, the action and state space, the rewards,
and, finally, the training method.
2.1. Background
2.1.1. Reinforcement Learning
A Reinforcement Learning (RL) model is typically formalized as a Markov Decision
Process (MDP) represented by the 5-tuple
M= (S
,
A
,
p
,
γ
,
R)
. In this tuple, S denotes the
state space (all possible states that the environment can be in at any given time), A is the
action space (all possible actions the agent can take in the environment), p is the environ-
ment dynamics function (rules or probabilities that define how the environment transitions
between states),
γ
is the discount factor (weighing future rewards relative to immediate
rewards), and R is the reward function (feedback mechanism that the environment uses to
indicate the success or failure of the agent’s actions) [
11
]. As an environment, we define the
world context in which our RL agents operate.
In our agent system, we use a centralized critic for the environment dynamics function
p and incorporate a discount factor
γ
of 0.99, as they are set by the ml agents package and
our configuration files. The observation space, action space, and rewards are discussed in
detail in the next sections.
Training in RL occurs through iterative interactions between the agent and the envi-
ronment. The agent learns to maximize cumulative rewards by selecting actions, receiving
feedback, and adjusting its policy, which is the strategy it uses. Each step provides rewards,
guiding the agent to refine its decision-making strategy. The process unfolds across multi-
ple episodes, which are sequences of interactions that start from an initial state and end
when a specific goal or condition is reached. Over time, the agent balances exploration
(trying new actions) and exploitation (leveraging learned strategies) to optimize long-term
outcomes [11].
It is important to note that multiple autonomous cooperating agents, collectively
referred to as a “multi-agent system”, operate within a shared environment. These agents
observe the environment using sensors and interact with it through actuators. While

Appl. Sci. 2024,14, 12068 4 of 20
pre-designed behaviors can be embedded in such agents, they often need to learn new
behaviors and actions online. This continuous learning process leads to improvements in
the performance of individual agents, as well as the overall multi-agent system [12].
We employed the MA-POCA (Multi-Agent Posthumous Credit Assignment) algorithm
from the ML-agents package [
13
]. The algorithm learns a centralized value function to
estimate the expected discounted returns of the group of agents and a centralized agent-
centric counterfactual baseline to facilitate credit assignment [
8
]. Essentially, this allows for
the agents that have escaped from the environment to continue training and learn quicker
and more effectively.
This was the main reason we chose MA-POCA against others like Independent Q-
Learning that can be limited to complicated and dynamic environments [
14
] or the Multi-
Agent Advantage Actor-Critic (MA-A2C) that struggles in synchronization across agents
and may struggle in high-dimensional environments [15].
2.1.2. Personality Traits
We have adopted the OCEAN Five personality characteristics model, which stands for
Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. This model
is one of the most widely recognized and utilized frameworks for personality assessment in
the field of psychology [
16
]. The OCEAN model provides a comprehensive and scientifically
grounded approach to understanding individual differences, making it a valuable tool for
analyzing behavior in various contexts, including gaming and human–agent interaction.
While newer models offer useful insights, we have chosen to implement the OCEAN
Five model due to its widespread acceptance, empirical support, and versatility. The OCEAN
model remains the most commonly used framework for personality assessment, offering
a reliable and consistent structure for studying a broad range of personality traits and
behaviors across diverse populations. Furthermore, its applicability and recognition make
it a practical choice for our work, ensuring that our findings align with established psy-
chological research and are easily interpretable by both researchers and practitioners in
the field.
2.2. Related Work
Previous research has investigated the development of adaptive agents within serious
game environments to address various domain-specific challenges. For instance, a notable
example is the multi-agent system integrated into the SIMFOR project, which is a serious
game designed for crisis management training and simulation [
17
]. The agents in the
SIMFOR system are built on the Belief–Desire–Intention (BDI) deliberation model, which
enables them to simulate complex decision-making processes. These agents can be cus-
tomized and configured to facilitate the construction of diverse crisis scenarios, thereby
supporting training exercises in emergency response and crisis management.
In contrast, our work focuses on a fundamentally different goal. The multi-agent
system we have developed is not constrained to a specific domain like crisis management
but is designed to establish a generalized and standard methodology of personality emula-
tion. Rather than contributing to the construction of specific training scenarios, our system
emphasizes the emulation of personality-driven behaviors and adaptive traits that can be
seamlessly applied to agents in both serious games and entertainment-oriented games.
A multi-agent system based on decision trees with data from traditional questionnaires
has been developed to assess personality types [
18
]. It helps determine a person’s dominant
personality, suggest job placements, or predict reactions based on specific traits.
The main difference between this approach and ours is that it relies on self-administered
questionnaires, which can introduce biases like self-reporting errors. In contrast, our sys-
tem uses in-game behavior data, minimizing these biases for a more objective assessment
of personality.
Similarly, previous studies have used game environments to simulate behavior mod-
eling [
9
], such as analyzing how a single agent’s movement in a simple room reflects its

Appl. Sci. 2024,14, 12068 5 of 20
openness personality trait. Similarly, [
5
] demonstrated how escape rooms can measure
personality through gameplay metrics and puzzles. In contrast, our work develops a multi-
agent system with a reward function methodology to assess team efficiency in a gamified
environment while exhibiting complex behaviors driven by multiple personality traits.
2.3. Game Mechanics and Environment
Our multi-agent system was implemented as a 3D environment in the Unity platform.
Focusing on creating an engaging yet accessible ER experience, we harnessed Unity’s assets
to construct a captivating setting.
Our designed environment consists of two buttons (on the walls) that, when activated,
unveil a key required for unlocking the final door to escape, as can be seen in Figure 2.
The team, consisting of 4 agents, must navigate through the pillars and past the columns to
seek out and press buttons.
Figure 2. Unity Environment implementation. The agents and the buttons are illustrated with boxes
in the left image, while the key and the agents pressing the buttons are shown with the arrows in the
right one.
With each new game, the positions of the buttons, keys, door, and pillars are dynam-
ically generated and positioned within the room at random intervals. Notably, the dy-
namism of our environment adds a layer of unpredictability to the gameplay and further
enhances the immersive experience, and it tests the adaptability of the agents. This ensures
that the agents are constantly faced with fresh challenges and opportunities for discov-
ery that demand quick thinking and strategic problem-solving skills. This dynamic and
ever-changing environment ensures that each gameplay session is unique and offers new
opportunities for the agents to learn.
2.4. Action Space
In this multi-agent environment, the action space available to each agent is deliberately
structured to encompass four distinct options, which are tailored to facilitate navigation
and interaction within the virtual landscape. These options are split into two categories:
movement and rotation. For movement, agents are equipped with the ability to advance or
retreat backward, enabling them to traverse the intricate terrain of the ER with precision
and agility. Similarly, rotation options afford agents the capability to pivot either left or
right, affording them the flexibility to survey their surroundings and strategize accordingly.
Each of these action choices is encoded as a Boolean variable, ensuring clarity and
efficiency in decision-making processes. This streamlined approach not only simplifies
the agents’ decision-making process but also enhances the overall responsiveness and
adaptability of the system. By empowering agents with a diverse array of movement
and rotation options, we aim to provide them with the requisite tools to navigate and
interact with the virtual environment in a manner that closely mirrors real-world behaviors
and capabilities.
In addition to the overall architecture of the multi-agent environment, our monitoring
system records and analyzes the unique features of each agent’s gaming style. This holistic
method enables us to gain deeper insights into the complexities of agent behavior, shining
light on crucial areas such as movement dynamics, interpersonal relationships, and involve-

Appl. Sci. 2024,14, 12068 6 of 20
ment with the environment, which can range from picking up keys and pressing buttons to
attempting to access the exit door before it is unlocked.
One key aspect is the agents’ movement patterns, providing valuable data on their
navigation strategies and spatial awareness within the virtual realm. By scrutinizing
their locomotion, we can discern tendencies and preferences that inform their decision-
making processes, offering valuable insights into their cognitive mapping abilities and
strategic maneuvering.
2.5. State Space
Before proceeding with any decision-making process, the agents observe and diligently
gather specific information about their surroundings. This includes information like the
location of buttons, keys, and doors, as well as their current state, such as whether they
are pressed, found, or unlocked. What’s more, when one agent picks up a key or pushes
a button, this information is immediately communicated to all the other agents through
their observations.
By leveraging these observations, agents can effectively assess their environment,
allowing them to make informed decisions and navigate through the virtual world with
precision and efficiency.
In our Unity implementation, agents utilize a sophisticated observation mechanism
centered around ray-cast observations. This advanced technique utilizes physics functions
to project a ray into the environment scene, providing agents with crucial insights upon a
successful intersection with a target object. This method returns a Boolean value for each
one of some predefined tags, creating a final vector.
So, we use ray-cast arrays that project a total of 15 rays, each of which checks for
specific tags, which include the key, buttons, doors, other agents, and pillars. Also, all the
agents share 3 Boolean variables as common knowledge, regarding the buttons (pressed or
not), the key (picked or not), and the door (unlocked or not). This results in a final state
space size of 78, which captures all the relevant details about the environment that the
agents need to be aware of.
2.6. Rewards
The optimal design of MDPs is often challenged by the issue of sparse (infrequently
or only upon achieving specific goals) reward [
11
] , which poses a significant obstacle to
the agent’s learning process. This challenge is also encountered in the context of an ER
setting, where the most significant rewarding events occur only when all agents manage to
escape, making it difficult for individual agents to identify the chain of events that led to
the successful outcome [11].
To address this issue, we propose a two-fold reward system. The first part of the
reward system is related to the ER game and the team’s ability to solve it. The second
part is composed of custom reward functions that are designed to promote specific agent
behaviors. These two types of rewards are monitored separately, with the former reflecting
the team’s success in escaping the room and the latter aimed at making individual agents
to exhibit specific behaviors while assessing the whole team’s performance.
Regarding the ER game, we have implemented appropriate reward functions for each
agent that reward them when they reach specific checkpoints in the room, such as picking
up a key. Furthermore, a final team reward is given when all agents successfully escape.
The rewards are meant to encourage the agents to strive toward their ultimate objective,
which is to escape the room, as opposed to just locating the key or opening the door as
soon as possible. The proposed reward system includes a time-based reward for each agent
when they find a key, unlock a door, or successfully escape. Additionally, a team reward of
multiplied time-based reward is awarded when all agents successfully escape.
As previously mentioned, agents are rewarded based on their behaviors, which are
modeled using advanced mathematical approaches inspired by the HiDAC crowd simula-
tion framework. HiDAC introduces mathematical formulas to represent specific behaviors

Appl. Sci. 2024,14, 12068 7 of 20
linked to personality traits. For example, panic behavior is associated with traits such as
conscientiousness and neuroticism. In HiDAC, personality traits are modeled as Gaussian
distributions, where each behavior is a function of the corresponding trait calculated using
a Gaussian value (ranging from 0 to 1).
While HiDAC formulas focus primarily on the Gaussian distribution itself, our imple-
mentation goes a step further by correlating these Gaussian values with gameplay actions.
Using HiDAC’s mathematical models as a foundation, we developed custom formulas
(Table 1) that map personality-based behaviors to specific in-game actions (e.g., running,
pushing other agents) or mechanics (e.g., collision detection).
We must note that in our implementation, the Gaussian values ranged from
−
1 to 1,
depending on if we want to train the agent with a low (
−
1 to 0) or high (0 to 1) trait. This
range allowed us to nullify the rewards of specific behaviors (when set the Gaussian to 0)
so that we could focus the train on specific traits.
Table 1. Traits to rewards based on actions and characteristics
Personality Trait Behaviors (Original) Reward (Custom)
Openness Train (Agent Characteristic −added new state)
Explore num of correct actions × 10
Conscientiousness Panic 0.3 × −2 × ΨC+ 2 if run & push
Impatience 0.3 × (1 – ΨC)
Right Preference If ΨC> 0 then ΨC× (times right/time) × 0.3
Extraversion Leadership 0.3 × mean speed × ΨE
Communication 1 if num of communication actions used ≥ΨE≥0.5
Impatience 0.3 × 2 × ΨE– 1 if ΨE> 0
Pushing 1 if num of push actions used ≥0.3 × ΨE≥0.5
Personal Space (Agent Characteristic −collider)
Walk speed Max walk speed + 1
Gesture Num of correct gestures × 10
Agreeableness Impatience 0.3 × (1 – ΨA) if run each step
Pushing 1 if num of push actions used ≥0.3 × (1 −ΨA)≥0.5
Right Preference 0.3 × (Times right/time) × ΨA
Wait Radius (Agent Characteristic −collider)
Wait Timer (Agent Characteristic −wait timer)
Neuroticism Leadership Mean speed × (1 −ΨN) × 0.5
Panic ΨN× 0.5 if run and push
So, the agents are placed in the game environment and, before their training, their
personalities are chosen, and the corresponding Gaussian is set. At the end of each episode,
the agent is rewarded based on the metrics of each action. So the process is the following:
• We select the trait we want to train the agent to emulate (e.g. conscientiousness);
•
We set the Gaussian to define the selected trait (Go = 0, Gc = 1, Ge = 0, Ga = 0, Gn = 0);
• During training at the end of each episode:
1.
We collect behavior action metrics ( panic behavior = number of push actions on
other agents);
2.
We reward the agent by multiplying the behavior metric with the corresponding
Gaussian (R = Panic behavior × (Gc + Gn)).
It is important to note that a single behavior can be associated with multiple personality
dimensions, as it may have both positive and negative effects. The relationships between
different types of behaviors and personality traits have been further explored and are
detailed in Table 2. For example, Leadership can be found only as a positive trait if a person
is positively conscientious. It can also be found as a negative attribute if he shows negative

Appl. Sci. 2024,14, 12068 8 of 20
agreeableness traits. Otherwise, if he has neurotic tendencies, the leadership behavior will
have both positive and negative influences.
Last but not least, the rewards were designed to balance task-oriented objectives (e.g.,
solving puzzles, escaping the room) with personality emulation and behaviors, since there
are three layers of rewards: the team reward (escaping), the game-oriented reward (pressing
a button or finding the key), and the behavior rewards that we analyzed.
Table 2. Behaviors, personality traits, and bibliography review.
Behaviors Traits Reference
Exploration Openness (+) [19]
Personal Space Openness (+) [20]
Panic Neuroticism (+),
Conscientiousness (−)[21]
Leadership
Extraversion (+),
Conscientiousness (+),
Neuroticism (−)
[22]
Impatience Agreeableness (−),
Conscientiousness (−)[23]
Training Openness (+) [24]
In conclusion, the rewards for MindEscape’s agents are based on the HiDAC frame-
work, with specific adaptations shown in Table 1. The agents, their actions, and custom
reward functions are built upon the core mechanics of HiDAC while introducing novel
methods to emulate the traits of the OCEAN Five personality model, as shown on Figure 3.
Figure 3. From HiDAC to Mindescape environment.
2.7. Training Methodology
Each agent in the team is trained on specific personality traits and behaviors, using
the prior reward model. The team’s objective is to ensure the successful escape of all agents
while striving to achieve the same tasks as before.
For the agent’s training, we tried different hyperparameter values, as seen in the
following Table 3. The best hyperparameter values were different for the simple agents
and the ones emulating personality traits, something that was expected since the second
agents learned to show more complex behaviors.

Appl. Sci. 2024,14, 12068 9 of 20
Table 3. Hyperparemets and best values.
Hyperparameters Values
batch size 64, 128 , 256, 516
buffer size 64,000, 128,000, 256,000, 516,000
learning rate 0.005
hidden units 512
number layers 2
Italics are for the best default agents and the underlined for the personality agents.
3. Training Results and Evaluation
To train the agents that imitate human behavior, we first trained a simple multi-agent
system to solve the ER. In all the following figures of this section (Figures 4–8), as well as in
the Appendix A, the X axis represents the training steps, and the Y axis shows the rewards
(or the time) for each step.
All the agent teams were trained for 25 million steps, which took around 6 h for each
team. For assessment purposes, we meticulously scrutinized group rewards, episode length,
and, in certain instances, behavior metrics to gauge their proficiency and effectiveness.
3.1. Default Team Rewards
The results of the best default multi-agent team (without imitating human behavior)
are shown in Figure 4. As we can observe, agents learned almost from the start to press
buttons and had some positive rewards. In contrast, the team learned to escape and solve
the room much later, almost at 3M steps, as can be seen by the first peaks of the green line.
This was due to the dynamically changing environment, and this performance was set as a
standard for the next agents, which was displayed in every figure in a light and shaded
way. We did not intend to compare the agents’ to the team rewards, but this is an indication
of when training is effective for the team.
Figure 4. Default (no personality) agents and agent teams training results.
Uniform Team Rewards
Following that, we implemented the reward functions as stated before, regarding the
behaviors, and trained agent teams with different kinds of traits. The teams consisted of
four agents, each with the same personality at first, and then combinations of different kinds
of traits. In the first four teams that were trained, all their members emulated behaviors
from only one of the available personalities: Openness, Non-openness, Conscientiousness,
Non-conscientiousness, Extrovert, Introvert, Agreeable, and Non-Agreeable, Neurotic, and

Appl. Sci. 2024,14, 12068 10 of 20
Non-neurotic. This way, we could set a baseline of how agents with the same personality
traits are performing.
In all the diagrams of this section, we observe the smoothed and actual rewards of
each team at an agent level and at a group level, as well as the default agent reward, to be
able to understand how the behaviors change the rewards and effectiveness.
In Figure 5, we observe the performance of agents with varying levels of Openness.
The team with high Openness scored lower rewards compared to the default agents. This
difference in performance is primarily due to the high Openness agents’ propensity to
explore more extensively. Their curiosity and desire for new experiences lead them to
spend more time investigating their environment, which, while potentially valuable for
gathering information, results in less efficient task completion.
The exploratory behavior of high-Openness agents can divert attention from immedi-
ate goals, causing delays and reducing overall effectiveness. This tendency to prioritize
exploration over direct action can lead to lower rewards, as these agents may overlook
simpler, more immediate solutions in favor of investigating less obvious possibilities.
Teams of agents with all the personalities were trained, and the diagrams with com-
plementary analysis can be found in the Appendix A. Table 4presents a comparison of
the teams, summarizing their mean reward, meantime, and success rate to provide an
overview of their overall effectiveness.
Figure 5. Openness (O) and Non-Openess (NO) agents and team training results.
We present the Openness agents diagrams as an indicative showcase, and we include
the diagrams of all the agents in the Appendix Afor clarity.
Table 4. Uniform personality teams and effectiveness.
Personality (+/−)Mean Reward 1Mean Escape Time 2Success Rate (All
Agents Escaped)
Default 1970 410 27%
Openness 1695/2154 505/411 37/20 %
Conscientiousness 1890/2005 446/430 29/21%
Extraversion 1776/1891 428/409 29/30%
Agreeableness 1819/2003 440/430 37/27%
Neuroticism 2077/1999 415/432 20/38%
1The final reward number based on agent actions. 2The mean play time in seconds.

Appl. Sci. 2024,14, 12068 11 of 20
We can observe from the Table 4, the influence of positive and negative traits on the
effectiveness of the teams. This showcases just how different the agents behaved in the
same environment.
3.2. Uniform Team Behaviors and Escape Time
After analyzing the rewards, it is essential to examine the behaviors exhibited by the
agents. In Figure 6, we observe the variations in pushing actions among agents, which were
influenced by their personality traits. Neurotic agents displayed a high frequency of actions,
closely mirroring the behavior of default agents. This similarity suggests that neurotic agents,
driven by their heightened sensitivity to stress and urgency, are more reactive and exhibit a
greater number of actions in an attempt to manage their environment and achieve their goals.
In contrast, calm, non-neurotic agents demonstrated significantly better cooperation.
Their lower frequency of actions indicates a more deliberate and thoughtful approach,
prioritizing coordination and strategic planning over immediate reactive behaviors. These
agents tend to focus on maintaining stability and harmony within the team, resulting in
more efficient and cohesive group dynamics. Their ability to remain composed under
pressure allows them to execute tasks with greater precision and teamwork, enhancing
overall performance.
Figure 6. Openness (O), Non-Openness (NO), Neurotic (N), Non-Neurotic (NN) Agents actions number.
Last but not least, one more element we need to look at is time. As seen in Figure 7,
the Openness team is the one that took the longest, something expected since they tend to
explore more. We can see that all the agents were generally slower than the default agents,
since they showed more complete behavior. We must also note that the Non-Openness
team was the quickest of all.

Appl. Sci. 2024,14, 12068 12 of 20
Figure 7. Agents mean play time (Openness (O), Non-Openness (NO), Conscientiousness (C), Non-
Conscientiousness (NC), Extrovert (E), Non-Extrovert (NE) Neurotic (N), Non-Neurotic (NN) agents).
3.3. Non-Uniform Team Rewards
The next step was to train agents so that each one had a different kind of personality
trait. We allocated the initial number of agents based on the 25 percent ratio of introverts to
extroverts in a community [25] to also the Neurotic and Non-Neurotic traits.
Based on this hypothesis, we conducted an experiment to observe the impact of intro-
ducing one Introvert into a team of three Extroverts (red) and one Extrovert into a team of
three Introverts (green), comparing their performance to the previous teams. As shown in
Figure 8, the team comprising three Introverts and one Extrovert (green) outperformed all
other configurations. The presence of the Extrovert agent proved beneficial, as they natu-
rally assumed a leadership role, effectively organizing and directing the Introverts, whose
preference for reflective and deliberate actions complemented the Extrovert’s initiative
and energy.
In contrast, the team with three Extroverts and one Introvert (red) performed similarly
to the all-introvert team but outperformed the all-extrovert team. The Extroverts’ com-
petitive nature caused inefficiencies, but the Introverts introduced a stabilizing element,
reducing competition. This allowed the team to work more cohesively and efficiently than
the all-extrovert team.
Overall, these findings highlight the importance of balancing team dynamics with
diverse personality traits. The Extrovert’s leadership can harness the strengths of Introverts,
while an Introvert’s presence in an Extrovert-dominated team can introduce a calming
influence, fostering a more harmonious and effective team environment.
In the same way as before, we showcase the results of other teams and with a variety
of personalities, while their results, diagrams, and complementary analysis can be found in
the Appendix A.
As we can see in Table 5, there was a significant change of the success rates of the
teams when there was more than one personality in the agents. For example, when we
included 1 Introvert, the team improved by 10% but on the other hand, when we introduced
1 Neurotic to a Non-Neurotic team, the effectiveness dropped by 5%.
Based on the results, it is safe to say that the personality traits of the agents of each team
do play a significant role in how the group operates, how efficient it is, and how quickly they
manage to escape. This means that the reward functions and how they are set can indeed
replicate and emulate how the traits and behaviors are exhibited in real life based on logic.

Appl. Sci. 2024,14, 12068 13 of 20
Figure 8. Extrovert (E) and Introvert (Non-Extrovers—NE), 3 Extroverts and 1 Introvert (EEEI), and
3 Introverts and 1 Extrover (IIIE) agents training results.
Table 5. Non-uniform personality teams and effectiveness.
Personality (+/−)Mean Reward 1Mean Escape Time 2Success Rate
(Difference 3)
3 Extroverts and 1
Introvert 1150 380 39% (+10%)
3 Introverts + 1
Extrovert 1318 403 48% (+18%)
3 Neurotic + 1 Non
Neurotic 1654 362 36% (+16%)
3 Non Neurotic + 1
Neurotic 1530 360 32% (−6%)
1
The final reward number based on agent actions.
2
The mean play time in seconds.
3
The difference (number in
parenthesis) with the corresponding uniform teams.
3.4. System Evaluation
Emulating human behavior in a game environment is a complex task that cannot be
accurately measured without involving a diverse set of human players. In some cases,
however, simulation results can provide valuable guidelines for predicting human be-
havior, especially when focusing on specific personality traits. This approach guided the
evaluation of our models, with the HiDAC crowd simulation serving as the baseline for
comparison [26].
The results of the training process support the initial hypothesis that the agents can
emulate human behaviors to some extent. This is evidenced by two key factors: the reward
values obtained during training and the visual inspection of the agents’ behaviors within
the game environment. Each agent exhibited a range of behaviors, shaped by the reward
functions, that mimic simplified versions of human actions. While individual agents
displayed different gameplay styles and actions, these variations reflect the interplay of
multiple personality traits, which manifest at varying levels depending on the mental state
of the player at any given moment.
Although personality is inherently complex, we were able to distill each trait into
simplified behavioral patterns, allowing us to create agents that simulate core aspects of a
specific trait. This simplification, however, may be considered a limitation of the work, as it
reduces the full depth of human personality into more basic representations.

Appl. Sci. 2024,14, 12068 14 of 20
For an effective evaluation of the agents’ behavior, it is essential to establish clear
behavioral ground truths.
As an initial benchmark, experimental results from Durupinar provide a useful point
of comparison. These results, summarized in Table 6, serve as a reference for assessing the
agents’ performance and offer insights into how well the modeled behaviors align with
expected personality traits. This shows that the agents indeed have similarities with the
human behaviors that they emulate.
Table 6. Tested scenarios used for result comparison and evaluation.
Traits HiDAC Behavior Our System Behavior
Openness
As Openness increases, the number of places they explore
increases, and thus, they leave the building later.
The mean escape time of the
Openness agents team is
increased (+25%).
Extroverts and Introverts
The Extroverts approach the attraction point in a shorter
time. In addition, when there are other agents blocking
their way, they tend to push them to reach their goal.
The Extroverts show better times
of escaping (+5%), as well as
higher push action metrics
(+35%)
Conscientiousness and
Agreeableness
The shortest time is achieved when Conscientiousness and
Agreeableness are the highest. The result is expected as
agreeable and conscientious individuals are more patient,
they do not push each other, and are always predictable,
as they prefer the right side to move on. Also, the longest
time is obtained when both values are minimal.
The agents that have high
Conscientiousness and
Agreeableness are quicker than
the ones with lower (+5% and
+4%, respectively) and lower
push actions (>−25%)
Neurotic
Agents that are Neurotic with less Conscientiousness tend
to panic more, pushing other agents forcing their way
through the crowd, and rushing to the door.
The Neurotic agents are quicker
to finish the room (+5%) though
less successful (−55%).
4. Discussion
Based on the findings, it can be confidently asserted that the individual personality
traits exhibited by the agents within each team have a substantial influence on the overall
dynamics of the group, including its operational efficiency and the speed with which they
achieve their objectives, such as escaping.
This underscores the critical importance of the reward functions employed and their
configuration, as they have the capacity to accurately mirror and replicate the diverse
range of traits and behaviors observed in real-life scenarios. In essence, the design and
implementation of these reward structures have the potential to faithfully emulate the
complexities of human traits and behaviors through logical frameworks.
In our analysis, we have observed intriguing trends regarding the performance of
agents based on their personality traits. Particularly noteworthy is the efficiency and
effectiveness exhibited by agreeable agents, who demonstrate a remarkable propensity for
cooperation and swift problem-solving. These individuals, characterized by their affable
and cooperative nature, seamlessly navigate through challenges, leveraging their strong
interpersonal skills to foster collaboration within the team dynamic.
On the other side, introverted agents, while equally effective in their problem-solving
capabilities, tend to approach tasks at a more deliberate pace. Their cautious and introspec-
tive nature often translates into a methodical approach to problem-solving, resulting in
slower but effective progress through the ER challenges.
Moreover, our findings reveal intriguing dynamics when teams are composed of
agents representing a spectrum of personality traits. In such scenarios, we have observed
challenges in collaboration, particularly during the initial stages of engagement. The diverse
array of personality traits within the team dynamic can lead to discrepancies in communi-
cation styles, decision-making processes, and conflict-resolution strategies, posing initial
obstacles to seamless cooperation.

Appl. Sci. 2024,14, 12068 15 of 20
These insights underscore the intricate interplay between individual personality traits
and collective team dynamics within the multi-agent environment. By understanding
and leveraging these nuances, we can tailor strategies and interventions to optimize team
performance, fostering a cohesive and synergistic approach to problem-solving within the
ER setting.
There is a foundational premise supporting the assertion that team efficiency within a
gamified environment, where agents must collaborate to achieve objectives, is intricately
linked to the individual behaviors and traits exhibited by each team member. The findings
from our study reinforce this notion, highlighting the significant impact of specific metrics
and rewards established within the system.
Furthermore, our findings offer valuable insights into the broader applicability of this
reward methodology beyond the confines of our specific multi-agent environment. The suc-
cess of our approach in shaping agent behaviors and fostering efficient collaboration serves
as a promising indication of its potential utility across a diverse array of gaming scenarios.
In essence, our study not only showcases the intricate relationship between individual
traits and team efficiency but also underscores the transformative potential of tailored
reward methodologies in shaping agent behaviors and optimizing performance across
various gaming contexts.
In addition, these results open the door for further investigation into how exter-
nal factors—such as task difficulty, time pressure, and environmental complexity—may
interact with agent personality traits to influence team dynamics. For instance, under time-
critical conditions, teams with dominant Agreeable or Conscientious agents may perform
better due to their cooperative and focused problem-solving tendencies, whereas under
exploratory tasks, more Open or Introvert agents might excel due to their propensity for
careful analysis and creativity.
Finally, the results not only showcase the relationship between individual traits and
team efficiency but also underscore the potential of custom made reward functions in
shaping agent behaviors and optimizing performance across various gaming contexts.
These findings show the way for further advancements in designing agents with personality
for interactive simulations, virtual training environments, and commercial games, where
emulating realistic team behaviors and personality can enhance immersion, engagement,
and problem-solving capabilities.
5. Conclusions
The present study introduced a multi-agent environment in the form of an ER, where
agents emulate the OCEAN Five Personality Traits characteristics based on custom reward
functions. The environment was designed in a way that permitted us to gather data
regarding the play style of the agents, as well as their interactions with the room and
between them.
The RL agents emulate human-like decision-making processes through several techni-
cal elements. The game environment is modeled as a high-dimensional space, capturing
items, player positions, and interactions. The action space defines how agents can move,
pick up objects, solve puzzles, and interact, with each action impacting the game state
based on the agent’s personality. We implemented a custom reward function designed
to balance task-oriented objectives, such as solving puzzles and escaping the room while
looking at personality emulation and behaviors. During the training process, the agents are
trained using advanced Deep RL algorithms like MA-POCA (Multi-Agent Posthumous
Credit Assignment), involving extensive game sessions to improve strategies through trial
and error. These elements together ensure that the RL agents provide a comprehensive and
dynamic understanding of personality traits within the game.
In summary, the multi-agent teams developed in this study successfully emulated
personality traits and exhibited distinct behaviors while solving tasks within an Escape
Room (ER) environment. During the training process, we collected sufficient data to analyze
and compare the efficiency and speed of each team based on the personality traits of their

Appl. Sci. 2024,14, 12068 16 of 20
respective members. The findings highlight that personality traits significantly influence
team dynamics, task-solving efficiency, and the speed at which specific objectives are
achieved. These results suggest that our approach can serve as a foundation for developing
agents capable of learning to solve tasks in other game environments emulating behaviors,
using these or similar reward functions. Furthermore, such agents could also be deployed
as Non-Playable Characters (NPCs) capable of exhibiting realistic, personality-driven
behaviors, enhancing immersion and player experience in interactive digital environments.
Moreover, the data generated can be used to train a Multi-Output Regression System
(MORS) using supervised machine learning algorithms as future work. This process will
use a MORS to assess personality traits by analyzing data and behavioral patterns exhibited
during gameplay by human players. Future work will investigate the implementation
of a MORS in live, interactive gaming environments, where the system assesses players’
behaviors and provides immediate personality insights.
Author Contributions: Conceptualization, G.L. and I.V.; Methodology, G.L. and I.V.; Software, G.L.;
Validation, G.L.; Resources, G.L.; Data curation, G.L.; Writing—original draft, G.L.; Writing—review
& editing, G.L. and I.V.; Visualization, G.L.; Supervision, I.V. All authors have read and agreed to the
published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The original contributions presented in this study are included in the
article. Further inquiries can be directed to the corresponding author.
Acknowledgments: This paper is a revised and expanded version of a paper entitled Machine
Learning Methods for Emulating Personality Traits in a Gamified Environment, which was presented
at Pireus, Greece, 11–13 September 2024.
Conflicts of Interest: The authors declare no conflicts of interest.
Abbreviations
The following abbreviations are used in this manuscript:
AI Artificial Intelligence
ER Escape Room
RL Reinforcement Learning
MDP Markovian Decision Process
MA-POCA Multi-Agent Posthumous Credit Assignment
OCEAN Openness, Consientiousness, Extraversion, Agreeableness, Neuroticism
MORS Multi-Output Regression System
MA-A2C Multi-Agent Advantage Actor-Critic
Appendix A. Training Results and Analysis
The Figure A1 showcases agents with varying levels of Conscientiousness. We observe
that both teams, those with high and low Conscientiousness, performed similarly to
the default agents. However, the Non-Conscientious team exhibited some significant
performance drops.

Appl. Sci. 2024,14, 12068 17 of 20
Figure A1. Conscientiousness (C) and Non-Conscientiousness (C) agents and team rewards.
In Figure A2, we can observe the performance of the Extrovert and Non-Extrovert
(Introvert) agents. Both teams outperformed the default agents, suggesting that the default
agents’ behavior lacks stability and efficiency. Extrovert agents often act independently,
focusing on individual tasks rather than cooperating with others. This tendency to work
alone can lead to a lack of coordination and synergy within the team. On the other hand,
Introvert agents tend to be less proactive and may hesitate to take initiative, which can
result in slower progress and less overall activity.
Figure A2. Extraversion (E) and Non Extraversion/Introversion (NE) agents and team rewards.
A similar behavior is shown in Figure A3 from the agents that have high or low
agreeableness. Non-agreeable agents tend to perform better since they proactively take the
initiative to achieve the goal. On the other hand, agreeable agents focus more on organizing
and coordinating with each other. This tendency to prioritize cooperation and harmony
can sometimes slow down their progress, as they spend more time on communication and
consensus-building rather than taking decisive action. As a result, non-agreeable agents,
who are more willing to act independently and assertively, often reach their objectives
more efficiently.

Appl. Sci. 2024,14, 12068 18 of 20
Figure A3. Agreeable (A) and Non-Agreeable (NA) agents and team rewards.
Last but not least, the Neurotic and Non-Neurotic agents, as shown in Figure A4,
tended to score better than the default agents. This difference in performance can be
attributed to their inherent behavioral tendencies in such scenarios. Neurotic agents, often
characterized by high emotional reactivity and sensitivity to stress, may excel due to their
heightened vigilance and urgency in completing tasks. Their propensity to anticipate
potential problems can lead to faster and more meticulous execution of their goals.
Figure A4. Neurotic (N) and Non-Neurotic(NN) agents and team rewards.
Conversely, the teams with one Non-Neurotic agent alongside three Neurotic agents
(red) outperformed all other configurations, as shown in Figure A5. This superior per-
formance can be attributed to the balancing influence of the Non-Neurotic agent, who
brings a sense of calmness and stability to the team. The Non-Neurotic agent’s emotional
resilience and steady demeanor help to mitigate the high reactivity and stress sensitivity
of the Neurotic agents. This calming effect reduces the overall anxiety within the team,
allowing the Neurotic agents to function more effectively and focus on their tasks without
becoming overwhelmed.

Appl. Sci. 2024,14, 12068 19 of 20
Figure A5. Only Neurotic agents (N) team, only Non-Neurotic agents (Nn) team, team with 3
Neurotic agents (NNNNn) and team with 3 Non-Neurotic agents (NnNnNnN) team rewards.
References
1. Bergner, R. What is personality? Two myths and a definition. New Ideas Psychol. 2020,57, 100759. [CrossRef]
2.
Martinez, K.; Menéndez-Menéndez, M.I.; Bustillo, A. A New Measure for Serious Games Evaluation: Gaming Educational
Balanced (GEB) Model. Appl. Sci. 2022,12, 11757. [CrossRef]
3.
Reisenzein, R.; Hildebrandt, A.; Weber, H. Personality and Emotion. In The Cambridge Handbook of Personality Psychology, 2nd
ed.; Matthews, G., Corr, P.J., Eds.; Cambridge Handbooks in Psychology; Cambridge University Press: Cambridge, UK, 2020;
pp. 81–100. [CrossRef]
4. Fotaris, P.; Mastoras, T. Escape Rooms for Learning: A Systematic Review. In Proceedings of the 13th European Conference on
Games Based Learning (ECGBL 2019), Odense, Denmark, 3–4 October 2019.
5.
Liapis, G.; Zacharia, K.; Rrasa, K.; Liapi, A.; Vlahavas, I. Modelling Core Personality Traits Behaviours in a Gamified Escape
Room Environment. Eur. Conf. Games Based Learn. 2022,16, 723–731. [CrossRef]
6.
Angelini, G. Big five model personality traits and job burnout: A systematic literature review. BMC Psychol. 2023,11, 49.
[CrossRef] [PubMed]
7.
Durupinar, F.; Pelechano, N.; Allbeck, J.; Gudukbay, U.; Badler, N. How the Ocean Personality Model Affects the Perception of
Crowds. IEEE Comput. Graph. Appl. 2011,31, 22–31. [CrossRef] [PubMed]
8.
Cohen, A.; Teng, E.; Berges, V.P.; Dong, R.P.; Henry, H.; Mattar, M.; Zook, A.; Ganguly, S. On the Use and Misuse of Absorbing
States in Multi-agent Reinforcement Learning. arXiv 2022, arXiv:2111.05992.
9.
Liapis, G.; Lazaridis, A.V.I. Escape Room Experience for Team Building Through Gamification Using Deep Reinforcement
Learning. In Proceedings of the 15th European Conference of Games Based Learning, Virtual, 23–24 September 2021.
10.
Liapis, G.; Vordou, A.V.I. Machine Learning Methods for Emulating Personality Traits in a Gamified Environment. In Proceedings
of the 13th Conference on Artificial Intelligence (SETN 2024), Piraeus, Greece, 11–13 September 2024.
11.
Sutton, R.S.; Barto, A.G. Reinforcement Learning: An Introduction, 2nd ed.; Adaptive Computation and Machine Learning Series;
The MIT Press: Cambridge, MA, USA, 2018.
12.
Silva, J.; Dutta, A. Multi-Agent Deep Reinforcement Learning for Multi-Robot Applications: A Survey. Sensors 2023,23, 3625.
[CrossRef] [PubMed]
13.
Unity ML-Agents Toolkit. 2021. Available online: https://github.com/Unity-Technologies/ml-agents (accessed on 21 December
2024).
14.
Foerster, J.; Nardelli, N.; Farquhar, G.; Afouras, T.; Torr, P.H.S.; Kohli, P.; Whiteson, S. Stabilising Experience Replay for Deep
Multi-Agent Reinforcement Learning. arXiv 2018, arXiv:1702.08887.
15.
He, K.; Doshi, P.; Banerjee, B. Latent Interactive A2C for Improved RL in Open Many-Agent Systems. arXiv 2023, arXiv:2305.05159.
[CrossRef]
16.
Jang, K.L.; Livesley, W.J.; Vernon, P.A. Heritability of the big five personality dimensions and their facets: A twin study. J. Personal.
1996,64, 577–591. [CrossRef] [PubMed]
17.
Oulhaci, M.; Tranvouez, E.; Fournier, S.; Espinasse, B. A MultiAgent Architecture for Collaborative Serious Game applied to
Crisis Management Training: Improving Adaptability of Non Played Characters. In Proceedings of the 7th European Conference
on Games Based Learning (ECGBL 2013), Porto, Portugal, 3–4 October 2013.

Appl. Sci. 2024,14, 12068 20 of 20
18.
Ramírez, M.R.; Moreno, H.B.R.; Rojas, E.M.; Hurtado, C.; Núñez, S.O.V. Multi-Agent System Model for Diagnosis of Personality
Types. In Proceedings of the Agents and Multi-Agent Systems: Technologies and Applications 2018, Gold Coast, QLD, Australia,
20–22 June 2018; Jezic, G., Chen-Burger, Y.H.J., Howlett, R.J., Jain, L.C., Vlacic, L., Šperka, R., Eds.; Springer: Cham, Switzerland,
2019; pp. 209–214.
19.
Abu Raya, M.; Ogunyemi, A.O.; Broder, J.; Carstensen, V.R.; Illanes-Manrique, M.; Rankin, K.P. The neurobiology of openness as
a personality trait. Front. Neurol. 2023,14, 1235345. [CrossRef] [PubMed]
20.
Nam, N.; Hang Nga, N. Influence of personality traits on creativity and innovative work behavior of employees. Probl. Perspect.
Manag. 2024,22, 389–398. [CrossRef]
21.
Javaras, K.N.; Schaefer, S.M.; van Reekum, C.M.; Lapate, R.C.; Greischar, L.L.; Bachhuber, D.R.; Love, G.D.; Ryff, C.D.; Davidson,
R.J. Conscientiousness predicts greater recovery from negative emotion. Emotion 2012,12, 875–881. [CrossRef] [PubMed]
22.
Li, W.; Zhang, H.; Zheng, Y. Personality and Leadership: A Critical Review and Future Research Agenda from a Dynamic
Perspective. In Oxford Research Encyclopedia of Business and Management; Oxford University Press: Oxford, UK, 2024.
23.
Jiang, N.; Shi, M.; Xiao, Y.; Shi, K.; Watson, B. Factors Affecting Pedestrian Crossing Behaviors at Signalized Crosswalks in Urban
Areas in Beijing and Singapore. In Proceedings of the ICTIS 2011: Multimodal Approach to Sustained Transportation System
Development: Information, Technology, Implementation, Wuhan, China, 30 June–2 July 2011. [CrossRef]
24.
Bergold, S.; Steinmayr, R. Personality and Intelligence Interact in the Prediction of Academic Achievement. J. Intell. 2018,6, 27.
[CrossRef] [PubMed]
25.
Maureen, I.; Imah, E.; Savira, S.; Anam, S.; Mael, M.; Hartanti, L. Innovation on Education and Social Sciences: Proceedings of the
International Joint Conference on Arts and Humanities (IJCAH 2021) October 2, 2021, Surabaya, Indonesia, 1st ed.; Routledge: London,
UK, 2022. [CrossRef]
26.
Xu, Z.; Bai, Y.; Zhang, B.; Li, D.; Fan, G. HAVEN: Hierarchical Cooperative Multi-Agent Reinforcement Learning with Dual
Coordination Mechanism. Proc. AAAI Conf. Artif. Intell. 2023,37, 11735–11743. [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.