Welcome to nature.com. We appreciate your visit. It seems that your browser may have limited support for CSS. For the optimal site experience, we recommend using an updated browser or disabling compatibility mode in Internet Explorer. To ensure continued support, we are presenting the site without styles and JavaScript.
We are sharing an unedited version of a manuscript to provide early access to its findings. Please note that the manuscript will undergo further editing before final publication. As a result, there may be errors in the content, and all legal disclaimers apply.
The integration of social and emotional intelligence is crucial for human cognition. However, existing artificial agent frameworks often separate these capabilities, hindering their ability to create genuine social interactions. In response to this, we introduce SELAgents (Social and Emotional Learning Agents), a pioneering framework that merges emotional processing, theory of mind, and social learning into a unified reinforcement learning structure.
SELAgents incorporates a three-dimensional emotional state space (Pleasure-Arousal-Dominance model), Bayesian belief networks for mental state inference, and game-theoretic social strategy selection. Through rigorous experiments involving diverse agent populations over multiple time steps, we uncovered notable enhancements over conventional reinforcement learning frameworks. Emotional intelligence scores increased by 49%, social coherence improved by 66%, and resource allocation efficiency reached 87% in comparison to baseline models.
The agents exhibited various emergent behaviors, such as emotional contagion effects and stable coalition formations, reflecting patterns akin to human social dynamics. Our findings highlighted the significant contribution of theory of mind capabilities towards agent performance, closely followed by emotional processing and social strategies.
We have made the complete implementation of SELAgents available as open-source software to support further research endeavors. It’s important to note that this study assumed perfect observability of emotional states, setting a performance benchmark. Extending the framework to settings with partial observability remains a noteworthy path for future exploration.
For additional details and access to the data and code used in the study, please visit the following link: https://github.com/nicolastorresr/SELAgents.
To foster reproducibility and stimulate further research in this field, we have openly shared our implementation on GitHub. For comprehensive references and related studies on emotional intelligence, social dynamics, and multi-agent systems, please refer to the citation list provided in the original content.
We extend our gratitude for the research support from ANID FONDEF IDeA ID25I10018 and the Department of Electronics, Universidad Técnica Federico Santa María, Chile.
Nicolás Torres
Corresponding Author
The article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, enabling non-commercial utilization, sharing, distribution, and reproduction with appropriate attribution and without alterations. Please consult the full license terms for complete details.
If you are interested in the AI and Robotics research field, sign up for the Nature Briefing: AI and Robotics newsletter to receive essential updates weekly.
About the Author
