A NEURAL NETWORK DYNAMIC SEQUENCER FOR DISTRIBUTED MISSION PLANNING AND CONTROL

FUTURE SPACE MISSIONS WILL INCREASINGLY BE UNDER JOINT MANAGEMENT OF DISTRIBUTED CONTROL CENTERS WHICH ARE RESPONSIBLE FOR DIFFERENTCOMPONENTS OF THE MISSION OBJECTIVE. THE COLLECTION OF REALTIME TASKS REQUIRED OF THESPACECRAFT WILL HAVE DIFFERENT PRIORITIES, RELEASE DATES, DEADLINES, INTERDEPENDENCIES, AND EVEN CONFLICTING DEMANDS ON INSTRUMENTS, COMMUNICATION CHANNELS, AND EXPENDABLE RESOURCES. WE PROPOSE HERE THE INVESTIGATION OF A NEURAL NETWORK-BASED SEQUENCER FOR THE GENERATION AND CONTINUOUS OPTIMIZATION OF TASK SCHEDULES WHICH, TAKING INTO ACCOUNT RESOURCE CONSTRAINTS, AND INTERNAL STATES AND EXTERNAL ENVIRONMENTAL DYNAMICS (BOTH SCHEDULED AND UNSCHEDULED), BALANCES THE REQUIREMENTS OF DISTRIBUTED CONTROL CENTERS. THE CENTRAL INNOVATION OF THE DYNAMIC SEQUENCER WILL BE AN OPTIMIZING SCHEDULING LOOP. THE HEART OF THIS LOOP IS A MODIFIED HOPFIELD NETWORK WHOSE MANY-TERMED ENERGY FUNCTION IS CONSTRUCTED TO REFLECT TASK PRIORITY, PROCESSING EXPENDITURE (TIME AND RESOURCES), COMMUNICATION REQUIREMENTS, RESOURCE CONSTRAINTS, AND TIME DEPENDENCIES. THESE TASK CHARACTERISTICS ARE DESCRIBED BY FEATURE VECTORS IN A MULTIDIMENSIONAL REPRESENTATION SPACE. IN THIS REPRESENTATION SCHEDULED EVENTS AND TASK INTER-DEPENDENCIES ARE INCORPORATED THROUGH (POSSIBLY VARIABLE) TASK DEADLINES AND RELEASE DATES. PARAMETERS OF THE ENERGY FUNCTION AND ACTIVITY PROPAGATION RULE IN THE MODIFIED HOPFIELD NETWORK ARE DYNAMICALLY CONTROLLED BY INFORMATION OF TASK QUEUES AND TASK ENVIRONMENT ON HAND. OPTIMIZED TASK ASSIGNMENT SEQUENCES WITHIN EVENT DRIVEN VARIABLE TIME WINDOWS ARE OBTAINED THROUGH EVOLUTION OF THE NETWORK.