Project.Report-CoRe123

Weinan Qian, Ningxin Pan, Feiyang Xie, Haochen Vector Zhao

Keywords: Emergent language; reinforcement learning; deep learning; psychology; cognitive reasoning; communication

Abstract: Verbal communication is a crucial skill for humans, enabling the exchange of intricate information efficiently. However, the process by which this communication protocol emerges remains a mystery. With the rapid development of Artificial Intelligence (AI) algorithms, it is now possible to design virtual agents and environments to study the development of emergent language. In this work, we introduced new tasks, environments, and agents to simulate the emergence of language in a multi-agent system. We explored both supervised learning and Reinforcement Learning (RL) methods to train the agents. While supervised learning demonstrated the potential to train a reasonable emergent language, the training process exhibited instability. Contrary to initial expectations, our experiments revealed that simple RL algorithms may not be sufficient for effective language emergence. We delved into the complexities and challenges encountered with Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) techniques, emphasizing the need for nuanced approaches to address the intricacies of linguistic development in autonomous systems. Our code is available at https://github.com/SouthwestWindQ/Keep_Talking_and_Nobody_Explodes.

Submission Number: 231