Exploring the Importance of Information Relevance, Ontology and Utilities for Scalable Multi-agent Coordination

In the process of decentralized team coordination, when cooperative agents cannot observe the complete state of the team and environment, communication is important for agents to share their knowledge and policies to achieve the common goals. In this paper, based on teamwork model, we model information sharing as an independent decision process to improve agents' joint actions and achieve higher team reward. However, precise calculation of information utilities in large teams is computationally hard. To make practical information decision, instead of precisely calculate the expected utility, we design agents that, similar to human beings, can share information by evaluating the relevance between pieces of information. The key is that agents can automatically infer the information relevance from semantic relationship of information based on ontology graph and agents' local knowledge. Therefore, when agents get more relevant information, the information will be used to update their local knowledge as well as the relevant measurements. It will be greatly helpful to make more precise information decision to get more related information so that their model will be reinforced. Our simulations show that by using evaluated information relevance and coordination relevance from given ontologies, agents can share information efficiently.