Space Applications for Distributed Constraint Reasoning

1. Space Applications forDistributed Constraint Reasoning Brad Clement, Tony Barrett Artificial Intelligence GroupJet Propulsion Laboratory California Institute of Technology brad.clement@jpl.nasa.gov http://ai.jpl.nasa.gov/

2. Outline • Applications • multi-spacecraft missions • “collaborative” mission planning • network scheduling • Current approaches • Challenges for DCR • Unsolicited opinions

3. Multi-Robot Control • Goal selection • Future commanding • Commanding now • mode estimation/diagnosis • Perception & actuation

4. Control of/by Humans • Goal selection • Future commanding • Commanding now • mode estimation/diagnosis • Perception & actuation

5. Analyst Planner Executive Control Distributed Constrained Optimization Optimize a function of variable assignments with both local and non-local constraints.

6. Space Applications Decentralize decision-making? • competing objectives (self-interest) • control is already distributed • communication constraints/costs • computation constraints • multiple rovers • spacecraft constellation • Earth orbiters • Mars network • DSN antenna allocation • mission planning • construction, repair • crew operations

7. Applications – Multiple Spacecraft Origins Program Over 40 multi-spacecraft missions proposed! • Autonomous single spacecraft missions have not yet reached maturity. • How can we cost-effectively manage multiple spacecraft? NMP Earth Observing System Sun-Earth Connections NMP Structure & Evolution of the Universe Mars Network

8. Signals from Celestial Sphere y  x Signals from Magnetosphere  t Applications – Multiple SpacecraftClassification of Phenomena(Underlying Scientific Questions) Five Classification Metrics • Signal Location • Where are the signals? • Signal Isolation • How close are distinct signals in phenomenon? • Information Integrity • How much noise is inherent in each signal? • Information Rate • How fast do the signals change? • Information Predictability • How predictable is the phenomenon?

9. Isolation & Integrity Rate & Predictability High High Noise Rate SingleSpacecraft SingleSpacecraft Low Low Low High Low High Resolution Need Predictability Applications – Multiple SpacecraftMultiple Platform Mission Types Signal Separation Signal Space Coverage Signal Combination

10. Cross-links GN&C GN&C GN&C GN&C GN&C GN&C GN&C GN&C GN&C Executive Executive Executive Executive Executive Executive Executive Planner Planner Planner Planner Planner Planner Planner Analyst Analyst Analyst Analyst Analyst Analyst Analyst Space Applications – ScienceHow to Distribute? Who gets which components?

11. GN&C GN&C GN&C Executive Executive Executive Planner Analyst Autonomous Signal Separation • Why many executives? • Each spacecraft can have local anomalies. • During an anomaly communications can be lost due to drift. • Why only one planner? • During normal operations the spacecraft are guaranteed to be able to communicate. • Since spacecraft join to make an observation, only one analyst is needed.

12. Autonomous Signal Space Coverage • Why many planners? • Cross-link is lost during normal operations, but spacecraft still have to manage local activities and respond to science events. • Why communicate at all? • The value of local measurements is enhanced when combined with data from others. Analysts must coordinate over collection. GN&C GN&C GN&C Executive Executive Executive Planner Planner Planner Analyst Analyst Analyst

13. GN&C GN&C GN&C Executive Executive Executive Planner Planner Planner Analyst Analyst Analyst Autonomous Signal/Mission Combination • How does this differ from signal space coverage? • Each entity has different capabilities • Sensors: radar, optical, IR... • Mobility: satellite, rover... • Communications abilities. • Each mission has its own motivations. • There is a competition where each mission wants to optimize its own objectives in isolation.

14. Commands (L-Band) { AFSCN } Payload Data Telemetry, QL Payload (S-Band) (X-band) Spacecraft GS (i.e., RSC) Payload GS (i.e., Datalynx, USN) Payload Downlink Requests Payload Data - FTP - Overnight (all) PTF Activity schedules Telemetry (ftp) Pager TS-21 Engr R/T MOC MOC ASPEN SCL Matlab Simulation Env SCL TT&C W/S Commanding SOH display Telemetry Pass Playback SOH display Trending Anom Res Payload Ops W/S Engineering Models Cmd Verification Mission Planning Data Center TT&C W/S TT&C W/S rescheduled activities Fight Dynamics newactivities rejectedactivities confirmation PPC Cluster Cmd Verification scheduleupdates MPW removedactivities local constraints Decentralize decision-making? • competing objectives (self-interest) • control is already distributed Space Applications – Mission Operations • communication constraints/costs • computation constraints Techsat-21 • multiple instruments on spacecraft contend for resources • multiple scientists may compete for one instrument (HST) • scientists work with operations staff to make sure goals can be safely achieved • plans must be validated (carefully simulated) • changes made by users in parallel invalidate validation

15. Applications - Deep Space Network (DSN)

16. Applications - Deep Space Network (DSN) Decentralize decision-making? • competing objectives (self-interest) • control is already distributed • communication constraints/costs • computation constraints • 56 missions • 12 antennas • different capabilities • shared equipment • geometric constraints • human operator constraints • some schedule as long as 10 years into future • some require schedule freeze 6 months out • complicated requirements originally from agreement with NASA with flexibility in antennas, timing, numbers of tracks, gaps, etc. • schedule centrally generated, meetings and horse trading to resolve conflicts • similar to coordination operations across missions

17. Applications – DSN Arrays Decentralize decision-making? • competing objectives (self-interest) • control is already distributed • communication constraints/costs • computation constraints • NASA may build 3600 10m weather-sensitive antennas • 1200 at each complex in groups of 100 spread over wide area • High automation requested—one operator for 100 or 1200 antennas • Spacecraft may use any number of antennas for varying QoS, and may need link carried across complexes • Only some subsets of antenna signals can be combined • depends on design of wiring/switching to combiners • combiners may be limited • Local response time should be minimized

18. Mars Network • Network traffic scheduled far in advance • Windows of comm availability • Need to react to unexpected events and reschedule • Missions must control own spacecraft • Comm affects resources that are needed for other operations • Continual negotiation MGS MEX Odyssey MER B MER A

19. How does DCR fit? • Goal selection • task allocation • Future commanding • meeting scheduling • Commanding now • mode estimation/diagnosis • Perception & actuation

20. Distributed Constraint Reasoningfor Planning & Scheduling • Allocating events/resources to time slots (meeting scheduling) • Hannebauer and Mueller, AAMAS 2001 • Maheswaran et al., AAMAS 2004 • Modi & Veloso, AAMAS 2005 • Coordinating plans by making coordination decisions variables • Cox et al., AAMAS 2005

21. = = Planner Planner Planner = Executive Executive Executive Shared Activity Coordination Shared activities implement team plans, joint actions, and shared states/resources

22. Shared Activity Coordination(SHAC, Clement & Barrett, 2003) • continual coordination algorithm • language for coordinating planning agents • framework for defining and implementing automated interactions between planning agents (a.k.a. coordination protocols/algorithms) • software • planner-independent interface • protocol class hierarchy • testbed for evaluating protocols

23. Shared Activity Model • parameters(string, integer, etc.) • constraints (e.g. agent4 allows start_time [0,20], [40,50]) • decompositions(shared subplans) • permissions- to modify parameters, move, add, delete, choose decomposition, constrain • roles- maps each agent to a local activity • protocols- defined for each role • change constraints • change permissions • change roles • includes adding/removing agents assigned to activity

24. Control Protocols for a Shared Activity • Chaos • A free-for-all among planners • Master/Slave • The master has permissions, slaves don’t • Round Robin • Master role passes round-robin among planners • Asynchronous Weak Commitment (AWC) • Neediest planner becomes master • Variations • how many planners share activity • use of constraints

25. Asynchronous Weak Commitment AWC::modifyPermissions() • if have highest priority • remove self’s modification permissions (add, move, delete) • else • give self modification permissions AWC::modifyConstraints() • if cannot resolve local conflicts and conflicts with constraints of higher ranking agents • set own rank to highest rank plus one • generate parameter constraints (no-good) describing locally consistent values

26. Experiments – Abstract Problem • joint measurements • capability matching • 3-9 spacecraft • 1-7 capabilities • 1-9 joint goals each requiring 1-4 of each capability

27. Experimental Results(Progress over cpu time) AWC Chaos - invalid solutions M/S - not complete RR Chaos M/S number of problems max cpu time (seconds)

28. Computing Consensus Windows 1 1 1 1 Agent C Agent C Agent A Agent B Agent A Agent B 2 2 highest rank decides voting or auction execute time Agent A Agent B Agent C consensus window

29. execute time Agent A Agent B Agent C consensus window Computing Consensus Windows 1 1 Agent C Agent A Agent B 2 2 voting or auction

30. Computing Consensus Windows 1 1 Agent C Agent A Agent B 2 2 voting or auction execute time Agent A Agent B votescollected Agent C consensus window

31. Computing Consensus Windows 1 1 Agent C Agent A Agent B 2 2 voting or auction execute time Agent A Agent B Agent C

32. Computing Consensus Windows 1 1 Agent C Agent A Agent B 2 2 voting or auction execute time Agent A Agent B Agent C consensus window

33. Causal Inconsistency Order of events • a is master and shares with (adds to roles) b • b receives add from a • a replaces b with c and makes c master • c receives add message making it master • c makes b master and removes self (deletes) • b receives add/master from c (before delete from a) • a receives update from c • b receives delete from a 2 add b 1 A SHAC protocol is proven sound if • the underlying planners are sound, • the protocol ensures that only one agent has permissions over any piece of information, and • it employs causally consistent communication 8 a delete 6 3 3 7 add/master add/master 4 5 update 5 c

34. Summary • Many space applications for distributed constraint reasoning • Many involve model-based causal systems • Need to map these systems to DCRs • how are CSPs mapped? • Need to handle • continuous variables (including cpu) • limited computation • not 1000 computers, but 2-10 • communication outages, unreliability, guarantees