The real-time execution performance agent an approach for balancing hard and soft real-time execution for space applications

Use of AI (Artificial Intelligence) algorithms such as adaptive planners, intelligent monitors, and data miners can help optimize overall return from space systems by providing adaptive operations that can exploit opportunities. Typically, space systems involve many hard real-time functions including: attitude, thermal, propulsion, and mechanism control; detector/sensor data stream processing; telemetry gathering and packetization; command handling; and many other periodic tasks which must be executed such that processing is completed by a periodic deadline. While there has been a concerted effort to design AI algorithms to have predictable execution requirements (e.g. anytime algorithms), most of these applications are relegated to running in a best effort fashion using slack time left over from the hard realtime periodic tasks which must be given higher priority to ensure safety and control. The problem with executing the AI algorithms in slack time is that this makes their execution performance impossible to predict. The alternative of requiring AI algorithms to be anytime algorithms so that they can be treated like a hard real-time task with a deterministic minimum response time can be prohibitive since such algorithms are hard to design and the minimum response may not provide much of an optimization. This paper describes an third alternative which provides an intelligent execution control mechanism, the EPA (Execution Performance Agent), that ensures execution of algorithms based on required reliability and confidence in meeting deadlines rather than priorities. The EPA provides predictable and safe execution of hard real-time safety critical and soft real-time mission optimizing tasks. By analogy, the EPA provides a balancing capability much like the everyday ability people have to walk without tripping while contemplating how to build a better career. It does this by executing tasks in specific execution reliability and confidence space and monitoring actual execution times to determine when resources must be adjusted. The EPA is currently being evaluated in a digital control and continuous video media testbed at the University of Colorado. Based upon testbed results, the EPA is also being considered for execution control of real-time operating system tasks including AI and digital control applications on a small spacecraft, Citizen Explorer, being built by the Colorado Space Grant College. The EPA was inspired by experience with a Space Grant Space Shuttle small payload which included control of three instruments and optimization of their operations using an adaptive planner and an intelligent monitoring system from the NASA Jet Propulsion Laboratory. The requirements for both hard real-time tasks and the use of AI applications on Citizen Explorer will be more demanding, and it is hoped the EPA can be shown to increase reliability and predictability of such systems. Details of the EPA mathematical formulation, the testbed implementation, performance results, and results of the analysis to determine if the EPA meets the Citizen Explorer requirements will be discussed in the paper.