Transformer-Based Approach for Social Norms Learning in Open Multi-Agent Systems

Open Multi-Agent Systems (OMAS) are societies composed of autonomous agents that behave with similar or different aims. Agents' cooperation, as a result, can be challenging when individual agents maintain their autonomy and nondesired situations may occur. To avoid that risk, norms are hardwired behavioral constraints used as a regulatory mechanism on autonomous agents when their observed behavior is considered to be abnormal. In contrast, the transformer model is a deep learning architecture known for its effectiveness in capturing long-range dependencies in sequential data. To that end, this paper proposes a novel transformer-based approach enabling agents to learn norms related to similar situations in different open environments. Based on our proposal, the agent will be context-aware and can detect and learn normative aspects by applying the outputs of norm-inference mechanisms. Also, the proposed approach is cognitively inspired in such a way as to capture long-distance dependencies and the correlation between observed agent actions and extensible using fine-tuning with new data while retaining its learned weights and parameters. The feasibility of our approach is illustrated through several scenarios.