90 likes | 200 Views
Commentary on: “FUSE Planning and Scheduling Under One Wheel Attitude Control”. Sandra Mau Robotics Institute, Carnegie Mellon University. Problem: Manual Labour. Additional constraints imposed by the failure of reaction wheels Much of the scheduling is done manually
E N D
Commentary on: “FUSE Planning and Scheduling Under One Wheel Attitude Control” Sandra Mau Robotics Institute, Carnegie Mellon University
Problem: Manual Labour • Additional constraints imposed by the failure of reaction wheels • Much of the scheduling is done manually • Very time consuming: 4-5 days to produce the next weekly plan • Seems like a full-time job!
Possible Alternative: Automation • Observation: • Scheduling is done offline • The task set is reasonably small (for the short term planning) • Problem is under constrained (many solutions) • It seems like there may be an optimal solution that can be found through combinatorial optimization • Difficulty: • Some constraints are global and order dependant • Many permutations possible and takes long to compute! • Solution: • Settle for a sub-optimal solution that is faster to calculate (as manual planners are currently doing.)
Possible Methods • Greedy scheduling using partially ordered list from Spike • Search algorithm (i.e. DFS, DFID, BFS, heuristic search) with initial state of empty schedule and final goal of satisfiable schedule. Can use a heuristic such as priority to get more important tasks scheduled first. • Both would require modification of short-term scheduler to work incrementally
Additional Questions • Spike is capable of taking into account quite a few of the constraints when it produces a schedule, including TA windows, beta and pole constraints, and hemisphere campaigns. What additional necessary considerations or constraints is it lacking? Can a modified short-term scheduler satisfy this? • Cost-Benefit - time required now, time it takes to develop automated software, FUSE lifespan?
Greedy Scheduling Return unscheduled observation to Spike list Spike’s 7-day list of partially ordered observations (START) NO Pop observation with desired +/- momentum from Spike list Modified short term scheduling software keeps track of previous slew and momentum and calculates how new observation would affect new state Slew & momentum still within bounds Compare to TA & momentum change plot tool to determine next desired momentum YES Initially empty 7-day MPS list of ordered observations (GOAL: FILL UP) Append schedulable observation to MPS list ITERATE
Search Tree • Similar steps to greedy but in a different order: Pop only feasible neighbours (tasks) at each iteration (checks before popping) • DFID with a priority heuristic might work well to find a satisfiable solution for this under constrained problem