Symbolic Knowledge Comparison: Metrics and Methodologies for Multi-Agent Systems

In multi-agent systems, understanding the similarities and differences in agents’ knowledge is essential for effective decision-making, coordination, and knowledge sharing. Current similarity metrics like cosine similarity, Jaccard similarity, and BERTScore are often too generic for comparing knowledge bases, overlooking critical aspects such as overlapping and fragmented boundaries, and varying domain densities. This paper introduces new specific similarity metrics for comparing knowledge bases, represented via symbolic knowledge. Our method compares local explanations of individual instances, preserving computational resources and providing a comprehensive evaluation of knowledge similarity. This approach addresses the limitations of existing metrics, enhancing the functionality and efficiency of multi-agent systems.