Evolving Agents: Interactive Simulation of Dynamic and Diverse Human Personalities (2404.02718v3)
Abstract: Human-like Agents with diverse and dynamic personalities could serve as an essential design probe in the process of user-centered design, thereby enabling designers to enhance the user experience of interactive applications. In this article, we introduce Evolving Agents, a novel agent architecture that consists of two systems: Personality and Behavior. The Personality system includes Cognition, Emotion, and Character Growth modules. The Behavior system comprises two modules: Planning and Action. We also build a simulation platform that enables agents to interact with the environment and other agents. Evolving Agents can simulate the human personality evolution process. Compared to its initial state, agents' personality and behavior patterns undergo believable development after several days of simulation. Agents reflect on their behavior to reason and develop new personality traits. These traits, in turn, generate new behavior patterns, forming a feedback loop-like personality evolution. Our experiment utilized a simulation platform with ten agents for evaluation. During the assessment, these agents experienced believable and inspirational personality evolution. Through ablation and control experiments, we demonstrated the effectiveness of agent personality evolution, and all of our agent architecture modules contribute to creating believable human-like agents with diverse and dynamic personalities. We also demonstrated through workshops how Evolving Agents could inspire designers.
- Christian Erik Kampmann. Feedback loop gains and system behavior (1996). System Dynamics Review, 28(4):370–395, 2012. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/sdr.1483.
- Generative Agents: Interactive Simulacra of Human Behavior | Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology.
- Humanoid Agents: Platform for Simulating Human-like Generative Agents, October 2023. arXiv:2310.05418 [cs].
- AgentSims: An Open-Source Sandbox for Large Language Model Evaluation, August 2023. arXiv:2308.04026 [cs].
- Voyager: An Open-Ended Embodied Agent with Large Language Models, October 2023. arXiv:2305.16291 [cs].
- MineDojo: Building Open-Ended Embodied Agents with Internet-Scale Knowledge. Advances in Neural Information Processing Systems, 35:18343–18362, December 2022.
- Ghost in the Minecraft: Generally Capable Agents for Open-World Environments via Large Language Models with Text-based Knowledge and Memory, June 2023. arXiv:2305.17144 [cs].
- LLM-empowered Chatbots for Psychiatrist and Patient Simulation: Application and Evaluation, May 2023. arXiv:2305.13614 [cs].
- Improving Factuality and Reasoning in Language Models through Multiagent Debate, May 2023. arXiv:2305.14325 [cs].
- Building Cooperative Embodied Agents Modularly with Large Language Models, February 2024. arXiv:2307.02485 [cs].
- Marvin Zuckerman. Psychobiology of Personality. Cambridge University Press, May 1991. Google-Books-ID: TA01Duy4RLwC.
- Cognitive complexity and dynamic personality in agent simulation. Computers in Human Behavior, 23(6):2983–2997, November 2007.
- Tuncer I. Ören. Dynamic Templates and Semantic Rules for Simulation Advisors and Certifiers. In Paul A. Fishwick and Richard B. Modjeski, editors, Knowledge-Based Simulation: Methodology and Application, pages 53–76. Springer, New York, NY, 1991.
- Sil Hamilton. Blind Judgement: Agent-Based Supreme Court Modelling With GPT, January 2023. arXiv:2301.05327 [cs].
- Math Agents: Computational Infrastructure, Mathematical Embedding, and Genomics, July 2023. arXiv:2307.02502 [cs, q-bio].
- The Rise and Potential of Large Language Model Based Agents: A Survey, September 2023. arXiv:2309.07864 [cs].
- A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis, February 2024. arXiv:2307.12856 [cs].
- Plan, Eliminate, and Track – Language Models are Good Teachers for Embodied Agents, May 2023. arXiv:2305.02412 [cs].
- Helping the Helper: Supporting Peer Counselors via AI-Empowered Practice and Feedback, May 2023. arXiv:2305.08982 [cs].
- Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI ’23, pages 1–21, New York, NY, USA, April 2023. Association for Computing Machinery.
- Interacting with Next-Phrase Suggestions: How Suggestion Systems Aid and Influence the Cognitive Processes of Writing. In Proceedings of the 28th International Conference on Intelligent User Interfaces, IUI ’23, pages 436–452, New York, NY, USA, March 2023. Association for Computing Machinery.
- 3DALL-E: Integrating Text-to-Image AI in 3D Design Workflows | Proceedings of the 2023 ACM Designing Interactive Systems Conference.
- Hey Dona! Can you help me with student course registration?, March 2023. arXiv:2303.13548 [cs].
- SAPIEN: Affective Virtual Agents Powered by Large Language Models*. In 2023 11th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW), pages 1–3, September 2023.
- Human-level play in the game of Diplomacy by combining language models with strategic reasoning | Science.
- CAMEL: Communicative Agents for ”Mind” Exploration of Large Scale Language Model Society. March 2023. Publisher: arXiv.
- Herding AI Cats: Lessons from Designing a Chatbot by Prompting GPT-3. In Proceedings of the 2023 ACM Designing Interactive Systems Conference, DIS ’23, pages 2206–2220, New York, NY, USA, July 2023. Association for Computing Machinery.
- BASES: Large-scale Web Search User Simulation with Large Language Model based Agents, February 2024. arXiv:2402.17505 [cs].
- Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST ’23, pages 1–22, New York, NY, USA, October 2023. Association for Computing Machinery.
- A Survey on Large Language Model based Autonomous Agents, March 2024. arXiv:2308.11432 [cs].
- Becoming Modern.
- ReAct: Synergizing Reasoning and Acting in Language Models, March 2023. arXiv:2210.03629 [cs].
- PaLM-E: An Embodied Multimodal Language Model, March 2023. arXiv:2303.03378 [cs].
- Personality Traits in Large Language Models, September 2023. arXiv:2307.00184 [cs].
- AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation, October 2023. arXiv:2308.08155 [cs].
- [2305.16291] Voyager: An Open-Ended Embodied Agent with Large Language Models.
- Training Socially Aligned Language Models on Simulated Social Interactions, October 2023. arXiv:2305.16960 [cs].
- The Embodied Mind, revised edition: Cognitive Science and Human Experience. MIT Press, January 2017. Google-Books-ID: bxLiDQAAQBAJ.
- CharacterChat: Learning towards Conversational AI with Personalized Social Support, August 2023. arXiv:2308.10278 [cs].
- Empathetic AI for Empowering Resilience in Games, February 2023. arXiv:2302.09070 [cs].
- The Big Five Trait taxonomy: History, measurement, and theoretical perspectives. In Handbook of personality: Theory and research, 2nd ed, pages 102–138. Guilford Press, New York, NY, US, 1999.
- Charles K. Coe. The MBTI: Potential Uses and Misuses in Personnel Administration. Public Personnel Management, 21(4):511–522, December 1992. Publisher: SAGE Publications Inc.
- Cognitive Architectures for Language Agents, March 2024. arXiv:2309.02427 [cs].
- Identifying and Manipulating the Personality Traits of Language Models, December 2022. arXiv:2212.10276 [cs].
- HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face. Advances in Neural Information Processing Systems, 36:38154–38180, December 2023.
- Computer says ”No”: The Case Against Empathetic Conversational AI, July 2023. arXiv:2212.10983 [cs].
- Least-to-Most Prompting Enables Complex Reasoning in Large Language Models, April 2023. arXiv:2205.10625 [cs].
- The ”What” and ”Why” of Goal Pursuits: Human Needs and the Self-Determination of Behavior: Psychological Inquiry: Vol 11, No 4.
- RobeRt Snowden. Oxford, UK: Oxford University Press, 2006 375 pages (£27.99 paperback).
- An Introduction to the Five‐Factor Model and Its Applications - McCrae - 1992 - Journal of Personality - Wiley Online Library.
- APA Dictionary of Psychology.
- Cognitive psychology, 7th ed. Cognitive psychology, 7th ed. Pearson Education New Zealand, Auckland, New Zealand, 2005. Pages: xxi, 602.
- Marvin Minsky. Society Of Mind. Simon and Schuster, March 1988. Google-Books-ID: bLDLllfRpdkC.
Summary
- The paper introduces an innovative architecture that enables continuous personality evolution in agents using coupled behavior and personality subsystems.
- The simulation platform built in Unity allows precise observation of agent interactions and the evolution of human-like traits in dynamic environments.
- Evaluation via the Big Five assessment and human reviews confirms that the system effectively mimics realistic behavioral changes.
Evolving Agents: Interactive Simulation of Dynamic and Diverse Human Personalities
Introduction
The paper "Evolving Agents: Interactive Simulation of Dynamic and Diverse Human Personalities" introduces a novel architecture for simulating human-like agents with dynamic personality traits and behaviors. This framework aims to address the limitations of static and homogeneous personalities in existing virtual agents by enabling continuous personality evolution through interaction with simulated environments and other agents. By integrating psychological theories of stimulus-response, the authors propose an architecture comprising two subsystems: Behavior and Personality, which cooperatively facilitate this evolution.
Figure 1: Evolving Agents architecture is inspired from psychology stimulus-response theory, encompasses two subsystems: Behavior and Personality.
Agent Architecture
Behavior System
The Behavior System consists of two core modules: Plan and Action. These modules facilitate personalized planning and social interactions based on current environmental contexts and character traits. The Plan module incorporates the agent's cognition and memory to devise plans that reflect its unique personality traits, leveraging a Characteristic Plan mechanism. Additionally, a Goal-Based Mechanism is introduced to map agent behaviors onto abstract goals, providing an analytical framework for personality-driven actions.
Figure 2: Each agent can create a schedule that aligns with its character. For example, the enthusiastic and open Isabella prefers to leave her own dorm and make social plans compared to the shy Benjamin.
The Action module enhances the depth of inter-agent communication by evolving conversation topics iteratively based on dialogue history. This facilitates a more natural progression in dialogue depth and content over time.
Figure 3: The agent discusses different topics with various partners, and the depth of each topic continuously increases.
Personality System
The Personality System comprises three modules: Emotion, Cognition, and Character Growth. Each module contributes to the dynamic update of an agent's character traits. The Emotion module is designed to capture the agent's instantaneous emotional state, influencing forthcoming plans through emotion-based plan evaluation.
Figure 4: Agents generate subjective descriptions that vary and align with their experiences.
Cognition is bifurcated into Memory and Insight, where Memory manages short-term behavioral data and summarizes it into long-term personality insights. The Insight module embodies holistic thinking processes to guide character evolution.
Figure 5: Insight enables agents to engage in personalized reflection on daily events.
Character Growth then updates the agent's personality structure, leveraging detailed algorithmic logic to reflect accumulated experiences and insights. The chain-like update process ensures each aspect of the personality adapts in concert.
Simulation Platform
The authors developed a simulation platform within a sandbox environment using the Unity engine. This platform allows users to observe agents' daily activities and thoughts, enhancing the realistic simulation of personality-driven behaviors.
Figure 6: In the sandbox world, users can observe the Agent's daily behavioral activities and inner thoughts.
A core innovation is the platform's three-level environment data structure, which enhances the efficiency of environmental interpretations by agents. This enables effective environment-agent interactions crucial for realistic personality simulations.
Evaluation and Results
To validate the architecture, the authors conducted simulations with three agents over several days, employing objective data verification and human evaluations. Objective metrics came from the Big Five personality traits assessment and behavioral similarity analysis. Results demonstrated meaningful personality evolution in agents assigned the full architecture, with distinct changes not observed in ablated control groups.
Figure 7: Results of the Big Five personality traits assessment.
Human evaluators confirmed the believability and perceptibility of the simulated personality differences and developments, thus substantiating the system's efficacy in emulating nuanced human-like traits.
Conclusion
The research presents Evolving Agents as a significant step toward realistic and dynamic human-like agent simulations. By integrating intricate behavioral and personality systems informed by psychological principles, the framework addresses limitations in previous agent simulations. Evolving Agents not only exhibit diverse, evolving personalities but also provide an interactive tool for design studies and psychological research.
The work demonstrates the potential of sophisticated simulated environments to yield insights into human behaviors and personalities, indicating promising future directions for the application of these systems in interdisciplinary research and design methodologies.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Collections
Sign up for free to add this paper to one or more collections.