Emergent behaviours considered harmful

Autonomous Agent Systems have proven difficult to instrument, monitor, control and test. Unit testing strategies are of little help, as they tend to stress functional aspects of a system rather than modal aspects and test at too low a level (the unit rather than the system). Further, emergent behaviours, indeterminacy, and the loose coupling of system elements often obscure the system's attributes. Absent specific inclusion in the system design of features that provide such specific capabilities, developers have little recourse to understand the systems they build. In this paper we describe ACME, a parallel agent system for testing used in UltraLog, a DARPA effort aimed at creating an ultra-survivable multi-agent system. We describe the motivation and architecture for a comprehensive testing system, and suggest a path for other MAS projects to improve testing and validation of their systems. Finally, we discuss the need for ad hoc logic changes in a running MAS, an instant messaging strategy for conversing with the system (including facilities for AliceBot and Jabber). We describe the use of the Ruby scripting language in the system and summarize the benefits and lessons learned.