Dual-task transformer for social norms learning in open multi-agent systems

Abstract

Open Multi-Agent Systems (OMAS) are societies composed of autonomous agents that operate with diverse or aligned objectives. To enhance cooperation among these independently functioning agents, researchers have adapted the concept of social norms from human societies as regulatory mechanisms. However, social norms often exist as unwritten, unspoken rules that can only be acquired through observation and inference, a significant challenge for autonomous agents. This paper proposes a novel Dual-task Transformer architecture that enables agents to learn social norms by observing interactions in their environment. Our approach leverages the transformer's capability to capture long-range dependencies between observed actions, performing two simultaneous tasks: contextually classifying action sequences to predict punishments and rewards, and generating structured representations of the inferred social norms. The model is trained on a modified version of the Moral-Stories dataset, transformed to represent sequences of agent actions and their normative outcomes. We integrate this inference capability with rule-based systems to assist agents in their decision-making processes. Experiments conducted within the MESA agent-based modeling framework demonstrate that agents equipped with our dual-task inference model rapidly internalize learned social norms and effectively regulate their behavior, showing a significant reduction in norm-violating actions and an increase in the adoption of norm-compliant behavior. Our approach offers a cognitively inspired mechanism for norm learning in OMAS that is both extensible through fine-tuning and capable of capturing complex dependencies in observed social interactions.