A Novel Reinforcement Learning Approach in Detecting Non-Collaborators for Trust Game with a Large Number of Participants
- Author: Ziyu Huang, Majoring in Data Science, Class of 2025, Duke Kunshan University
- Instructor: Professor Luyao Zhang, Duke Kunshan University
- Disclaimer: This is a submission for the Final Project in COMSCI/ECON 206 Computational Microeconomics, Spring 2024 Term (Seven Week - Fourth), taught by Professor Luyao Zhang at Duke Kunshan University.
- Acknowledgements: I am deeply grateful to my instructor, Luyao Zhang, for her exceptional guidance during the development of this computational economics paper. Additionally, I want to thank my colleagues for their support and collaboration, and my classmates for their stimulating discussions and valuable contributions.
This repository serves as a comprehensive collection for the course: Computational Microeconomics, covering groundbreaking topics in the field. It features a directory titled "CSECON" that investigates the dynamic realm of computational economics, spotlighting the confluence of human elements, AI, and computational technologies. Another section, "Advanced CSECON," scrutinizes the constraints of specific computational tools utilized in game theory experiments and critiques established studies on federated learning. Furthermore, the directory "proposal" presents a comparative study on the gaming behaviors of generative AI and reinforcement learning models, aiming to fill the voids observed in current research.
Hi all, I'm Ziyu Huang from China, junior student in Data Science. I've always thought economics to be an interesting subject, as I view it, its essential focus is on every individual in our society, endeavoring to accurately depict how they act and make decisions, particularly in the realms of trade and collaboration. This stands quite apart from Computer Science's pure mathematics and technical engineering focus. However, there are lots of CS applications can provide precise tools for quantitative analysis and forecasting for economics. I think this connection is my primary motivation for enrolling in this course. I expect this course will provide me with more interdisciplinary background, as well as introducing me to some new research methodologies or algorithms.
In this course, I leaned how advanced technologies like Generative AI (GAI), game theory, and multi-agent reinforcement learning (MARL) can address social and economic challenges. I have learn to model complex human behaviors and strategic interactions, enhancing problem-solving strategies in real-world scenarios. The magic of interdisciplinary research lies in its capacity to integrate diverse fields—merging AI with social sciences enriches our tools and approaches, leading to innovative, effective solutions. By understanding and applying these interdisciplinary methods, I not only gain a deeper insight into machine learning's potential but also its practical implications for societal benefits.
In this course, I have acquired numerous valuable tools, such as GitHub and Overleaf, which will greatly facilitate future collaborations with others. Learning to use these tools proficiently not only enhances my technical skill set but also prepares me to engage effectively in diverse team environments and multi-disciplinary projects, ensuring smooth and productive collaboration.
In terms of my personal views on the interdisciplinary research of computer science and economics, the study of generative AI is indeed a current popular area, and I firmly believe that this technology will eventually achieve recognition through awards like the Nobel Prize or the Turing Award. Yet evidently, it should not remain solely within the scientific domain. I assert that there is significant potential for a deep investigation into its implications for human society. I have observed that many papers in the field of economics have begun to explore how generative AI can imitate human decision-making processes and contribute to the advancement of game theory-related research. Nevertheless, there is currently a scant focus on how generative AI could revolutionize human cognitive processes and aid in a more profound self-comprehension.
My inspiration came from a paper by Takeshi and colleagues titled "Large Language Models are Zero-Shot Reasoners.". They discuss "chain of thought," a model of multi-step logical reasoning, which they used to assess the efficacy of ChatGPT. This sparked an epiphany for me due to a sudden connection I made with my own usage of ChatGPT. I had once envisaged it as a solution for all problems encountered in daily life, but my attempts ultimately met with failure. Upon reflecting on these unsuccessful endeavors, I recognized that my thought process is often nonlinear and that while analyzing issues, I could be influenced by many irrelevant factors that obscure the fundamental problem at hand. This bottleneck could be one of the reasons for ChatGPT's less-than-universal acceptance and might also explain the current surge in people investing efforts into the development of ChatGPT templates.