Conventions can be defined as recurrent behaviour patterns of human communities that increase the predictability of interaction outcomes. In the Artificial Intelligence context, conventions can coordinate agents’ actions in multiagent systems and simplify the agents’ decision-making machinery. In general, not all agents are necessarily willing or capable of adhering to the same conventions. However, in a cooperative multiagent system, we may assume that the agents will agree to adhere to the same conventions to improve their collective performance.
Previous research has demonstrated that conventions considerably enhance a multiagent system’s overall performance. Conventions are frequently hand-coded in those previous works, representing a costly effort for engineers. Any modification to the system’s conventions often requires manual checks of their soundness. Although previous work has introduced conventions into multiagent systems in an automated style, to our knowledge, no attempt has been made to develop an automatic NLP pipeline to embed conventions into a system. This research project aims to construct an architecture primarily concerned with processing natural language conventions. We believe that it will facilitate the creation of convention-based multiagent systems and give users control over multiagent behaviour. To illustrate the components of the architecture, we will use the board game Hanabi. Nonetheless, we will also attempt to present the architecture, particularly the natural language processing component (which contains a large language model), in as general terms as possible. Our vision is to “program” agents by declaring in natural language the game rules and the strategic behaviour that the agent should show.
Watch a video about Shuxian’s project: