Outriggers and Training Wheels for Cooperating Systems

When we build systems to operate in hazardous or remote environments, and especially when we expect them to cooperate with others in support of a goal, we rely on them to operate as correctly as possible in the (unpredicted, frequently unpredictable) situations they encounter. But the environment does whatever it does; we have essentially no control and only limited knowledge of what it does, and only the most meager notion of what it will do. In this paper, we describe an architecture for these component systems that we think will be better suited to the vagaries of environmental behavior than others. We advocate a collection of subsidiary systems to operate in parallel with the main system, to act as "outriggers" for unexpected environmental behaviors or system failures, or as "training wheels" during development. We describe our initial notions of how they relate to the original system and how to implement them.