Justification-based Multiagent Learning

Committees of classifiers with learning capabilities have good performance in a variety of domains. We focus on committees of agents with learning capabilities where no agent is omniscient but has a local, limited, individual view of data. In this framework, a major issue is how to integrate the individual results in an overall result—usually a voting mechanism is used. We propose a setting where agents can express a symbolic justification of their individual results. Justifications can then be examined by other agents and accepted or found wanting. We propose a specific interaction protocol that supports revision of justifications created by different agents. Finally, the opinions of individual agents are aggregated into a global outcome using a weighted voting scheme.