Verification of multiple agent knowledge-based systems

This paper's purpose is extending single agent verification tests to systems with multiple agent knowledge bases. Particularly, this paper identifies potentially anomalous situations occurring between agent knowledge bases. For example, consider one agent with rule "if A then B" and another agent with rule "if A then C," so agents have a "consequence conflict" for the same condition. In this setting, agents are constantly at odds, unable to reconcile differences. Alternatively, the following rules are found in one agent (if A then B, "if C then A"), while another agent has rule ("if B then C"). With two interacting rule bases, a dialogue starting with "A" cycles indefinitely, A --> B --> C --> A, depending on relationships between these two agents. Additionally, with traditional problems, this paper also identifies new issues, like agent "isolation" property. One potential approach to identifying anomalies in multiple agent systems is comparing each subset of agents' knowledge base determining existence verification issues. Where agents' number is small this approach is feasible. However, for even medium size systems this approach explodes computationally. As a result, alternative approaches are needed. This paper finds many multiple agent verification tests are conducted on metarule set generated from all rules contained in each of agents' knowledge bases minimizing computational effort and minimizing need for new verification concepts development. A wide range of "anomalies" are identified for multiple agent systems. Anomalies are system properties not necessarily incorrect. An important part of system structure is anomalous knowledge. For example, multiple agent systems employ negotiations and other devices facilitating agent interaction, mitigating anomalies. However, anomalies signal a system error. Ultimately, determination to dispostion of the anomaly is likely the responsibility of the developer. (C) 2001 John Wiley & Sons, Inc.