Abstract Infrastructure for Real Systems: Reflection and Autonomy in Real Time

CARS (Computational Architecture for Reflective Systems) is a low-cost test bed for studying self-organization and real-time distributed behavior, using cars with on-board computers as autonomous agents, in an uncontrolled and largely unpredictable environment. This paper describes the software infrastructure for CARS, based on our Wrapping approach to knowledge-based integration. It allows us to share code between simulations for algorithm development and instrumented experiments with the real cars in a real environment. It also allows us to use many computational resources during algorithm development, and then to ``compile-out'' all resources that will not be needed, and all decision processes that have only one choice, in a given real environment. The instrumented experiment is run in parallel with the simulation, and the differences can be used to adjust the models. We describe the autonomic agent infrastructure, i.e., the ``enabling software'' processes: health and status, local activity maintenance, and fault management. These processes can be very resource-hungry in any agent, and our use of simulations allows us to study trade-offs directly between safety and capability in the agents, to tune the trade-off at deployment time, based on what we know or expect of the environment, and to monitor and change those assumptions when necessary.