Traditional personality assessment relies heavily on self-report. Game-based AI assessment offers a more behavioural approach, but it creates a difficult research challenge: reliable models require labelled gameplay data, and collecting large, diverse datasets from human participants can be expensive and difficult to scale.
Our paper, “Multi-Agent System for Emulating Personality Traits Using Deep Reinforcement Learning,” published in Applied Sciences, proposes a solution based on deep reinforcement learning and multi-agent game environments. The paper was published on December 23, 2024, making December 24 an appropriate date for a research-focused blog post.mdpi
Traditional personality questionnaires provide labelled information because participants answer questions that are mapped to established personality dimensions.
However, game-based assessment works differently. It observes actions such as exploration, planning, cooperation, risk-taking, and persistence.
To train AI models to connect those actions with personality traits, researchers need substantial behavioural datasets with reliable labels. Collecting such data from human players at scale presents several challenges:
Our research explores how AI agents can help address this data problem.
The proposed system uses deep reinforcement learning to train agents whose behaviour is guided by the Big Five personality model, also known as OCEAN:
The objective is not to claim that artificial agents possess human personalities. Instead, we assign personality-related behavioural tendencies to agents and study how those tendencies affect their decisions and interactions.
An agent designed to express higher Openness may explore alternative paths and unfamiliar areas. An agent aligned with Conscientiousness may plan carefully and persist with a task. Other configurations may influence cooperation, risk management, conflict avoidance, or responses to uncertainty.
These behaviours create synthetic but controlled examples of trait-related gameplay.
In reinforcement learning, agents learn by receiving feedback from their actions.
Our framework defines reward functions based on OCEAN traits and the HiDAC behavioural system. These trait-aligned rewards encourage agents to develop patterns of action associated with particular personality profiles.
For example, rewards may encourage:
HiDAC provides a behavioural foundation for modelling variation between agents, while reinforcement learning allows those agents to learn how to act within the game environment. Related HiDAC research connects psychological and behavioural parameters with simulated agent behaviour.roderic.uv+1
The system uses Multi-Agent Posthumous Credit Assignment (MA-POCA) to support cooperative learning.
In a multi-agent environment, the team may receive a shared outcome even though each agent contributes differently. This creates a credit-assignment problem: how can the system determine which actions helped the team succeed?
MA-POCA is designed for cooperative multi-agent learning. It helps agents learn from shared outcomes and coordinate their behaviour toward a common objective.
This is particularly useful for studying personality in teams. Individual tendencies do not exist in isolation. The behaviour of one agent affects the decisions and performance of the others.
A highly exploratory agent may discover new information. A more conscientious agent may organise the group’s actions. A cooperative agent may support coordination, while a risk-averse agent may prevent the team from making a dangerous decision.
The result is a richer environment for studying how personality combinations influence collective performance.
The trained agents generate gameplay trajectories that reflect different personality configurations.
This produces synthetic behavioural data that can be used to:
The advantage of synthetic data is control and scalability. Researchers can create repeated scenarios, vary personality configurations, and test specific hypotheses without requiring a new human study for every experiment.
Synthetic data does not replace human data. It should be used as a complementary resource, particularly during model development and early experimentation.
This research contributes to the scientific foundation behind MindEscape and SeriousMind’s broader play-to-reveal platform.
MindEscape analyses how human players explore, solve puzzles, manage time, make decisions, and respond to pressure in immersive digital escape rooms. To develop reliable assessment models, researchers need to understand how different behavioural patterns may relate to personality traits.
Personality-emulating agents provide a controlled computational environment for studying those relationships.
They can help researchers ask:
These questions connect AI research with practical applications in personality assessment, recruitment, HR, team development, and serious-game design.
One of the most important practical questions is whether certain personality combinations lead to better team outcomes.
A team of highly exploratory agents may discover more possibilities but struggle to converge on a solution. A team dominated by highly conscientious agents may execute reliably but explore fewer alternatives. A balanced team may combine curiosity, planning, communication, cooperation, and persistence.
These examples are hypotheses, not universal rules.
Team performance depends on the task, the environment, the goals, and the interaction between individual behaviours. That is precisely why multi-agent simulations are valuable: they allow researchers to test different combinations systematically.
The findings may eventually inform how game-based assessments examine team fit and collaboration. They should not be used to make simplistic assumptions about individuals or to assign people to teams based on a single personality profile.
AI-generated agents can help build scalable behavioural baselines, but real-world assessment requires careful validation.
Human gameplay data remains essential for evaluating whether the behaviours generated by agents correspond meaningfully to human personality-related behaviour. Models must also be tested across different populations, game experiences, cultural contexts, and accessibility conditions.
At SeriousMind, our goal is to use AI to support better understanding—not to reduce people to fixed labels.
The research provides tools for exploring personality-coded behaviour in controlled environments. SeriousMind’s products translate those ideas into interactive experiences for individuals and organisations.
Together, the research and products form a bridge between:
The Applied Sciences paper addresses a central challenge in AI-powered personality assessment: how to develop scalable, trait-aware behavioural data without relying exclusively on costly human data collection.
By combining OCEAN-based reward functions, HiDAC behavioural modelling, DRL agents, and cooperative multi-agent learning, the research creates a framework for studying personality in dynamic game environments.
This work supports the broader vision behind SeriousMind: using serious games and artificial intelligence to understand how people think, act, collaborate, and respond to challenges.
The future of personality assessment will require both human insight and computational methods.
AI agents cannot replace human participants.
But they can help us build better questions, test stronger models, and understand behavioural patterns at a scale that would otherwise be difficult to achieve.
© 2026 SeriousMind. All Rights Reserved. Privacy Policy