IMPACT Multi-Agent Planning Research Dana Nau University of Maryland 1 Motivation: Noncombatant Evacuation Operations (NEOs) Goal: assist DOS to evacuate people whose lives are in danger noncombatants nonessential military personnel host-nation citizens third country nationals Characteristics: Joint task force – geographically distributed – often multinational Uncertainty; complexity (200+ tasks) US Ambassador is senior authority 1999 IMPACT Workshop 2 Difficulties with NEOs Multi-agent planning Supplies are not centralized Planning is the responsibility of the geographic commanders Resources: Doctrine, Exercises, DOS, EAP, etc. Information about supplies is not centralized either Potential for conflicts among simultaneous operations Example (not a NEO): moving Apache helicopters from Ramstein to Albania during the Kosovo crisis 1999 IMPACT Workshop 3 Key Technologies Access to distributed,heterogeneous data sources Seamless interoperability between different software capabilities Ability to coordinate multiple agents Scalable, high performance planning systems create plans interact with the above data and software sources Already being developed in the IMPACT project To be added to IMPACT 1999 IMPACT Workshop 4 What Planning Is Generate sequences of actions to perform tasks and achieve objectives Driving force The need for ways to aid human planning Example application areas military operations and logistics design and manufacturing space exploration 1999 IMPACT Workshop 5 Limitations of Classical Planning Classical planning theory: Either purely symbolic (AI planning) or purely numeric (OR planning) Single agent (the planner) Perfect information No interaction with users What’s needed in practice: Mixed symbolic & numeric computation Multiple agents Imperfect information, external information sources Human user in control of planning The above limitations may be overcome using a mix of IMPACT and scalable planning algorithms 1999 IMPACT Workshop 6 Approach Extend IMPACT to include (1) planning actions (2) temporal constraints for plans (3) complex planning tasks 1999 IMPACT Workshop 7 (1) Planning Actions Extend IMPACT rules so that heads of rules contain planning actions Example: if an item’s current inventory level has dropped below its restocking threshold, then create a plan to restock it at its stock level, by a given date, and within the given budget DO build-plan(restock(‘widget-25’,X.rstocklev,8/30/99,100k)) <= in(X,oracle:select(‘inventory”,item,=,‘widget-25’)) & <(X.qty,X.threshold). Problem: semantics of agent programs forces all executed actions to have a well-defined add/del list build-plan(…) may or may not be achievable, so we don’t know what the add/del list should be 1999 IMPACT Workshop 8 Planning Actions (Contd.) Extend IMPACT rule syntax to allow heads of rules to: contain a planning action and support actions whose outcomes may involve “conditional” add/delete lists of the form “fail” or “add/delete <specified lists>”. Extend the IMPACT implementation to support invocation of such planning actions 1999 IMPACT Workshop 9 (2) Temporal Constraints for Plans A tcass (temporallyconstrained action status set) is a set of action status atoms with temporal constraints on the actions Example: < {a1,a2,a3}, { st(a1) = st(a2), et(a1) < et(a2), st(a3) - st(a2) < 5, et(a3) - et(a2) < 3 } > where st(a) denotes a’s start time et(a) denotes a’s end time Increase the expressive power of agent rules by allowing rules of the form: Op tcass <= Body Also allow replacing tcass by a call to a planner which generates a tcass as its output We plan to refine the syntax and study semantics of such programs sound/complete status-set computation algorithms implementation techniques and experiments applications such as NEO planning 1999 IMPACT Workshop 1 (3) Complex Tasks Extend agent program rules so that the head of an agent program rule can have the form Op htn <= Body where htn is either an HTN (a hierarchical task network) – (see next slide) a call to an HTN generation program – e.g., the SHOP planning system (described later) We plan to refine the syntax and study the semantics computation algorithms for feasible status sets implementation and experiments applications 1999 IMPACT Workshop 1 HTN Planning: An Example Select Helicopter Launching Base Establish Base within Flying Distance alternative Launch from methods Carrier Battle Transport helicopters available (H) Security force available (F) Group Transport helicopters available (H) Helicopters have air refuel. capability (H) Select Helicopter Launching Base Select possible area (A) Transport sec. force (F,A,H) Embark sec. force (F,H) Fly(H,A) Disembark (F,H,A) Position security force (F,A) Transport fuel to (A) ... Decompose tasks into (more tactical) subtasks Consider restrictions (e.g., transport helicopters available) Resolve interactions (e.g., deploy security force first) If necessary, backtrack and try other methods 1999 IMPACT Workshop 1 Leverage (1) SHOP - Simple Hierarchical Ordered Planner New HTN planning system [Nau et al., IJCAI-99] Outgrowth of some of the ideas explored in the Bridge Baron Sound and complete over a large class of planning problems Much more expressivity than most other planning systems – Mixed symbolic & numeric computations – External information sources Solves standard benchmark problems orders of magnitude faster than other domain-independent planning systems Complete implementation in Common Lisp – Available via FTP; downloaded by dozens of researchers Implementation in Java underway 1999 IMPACT Workshop 1 Leverage (2) HICAP: Non-Combatant Evacuation Planning Joint ongoing work with the Naval Research Laboratory Combines SHOP with case-based reasoning – Makes use of military doctrine and previous successful plans Nominated for best-paper award at ICCBR-99 Héctor Muñoz will demo HICAP during the demo session 1999 IMPACT Workshop 1 Multi-Agent Planning We intend to do the following Incorporate the extensions mentioned earlier – planning actions – temporal constraints for plans – complex planning tasks Develop protocols by which multiple agents may coordinate planning activities with one another Derive results showing that (under certain conditions to be determined), such protocols guarantee convergence on a plan – This will ensure termination within predictable running times Develop applications – TBD, but a likely possibility is multi-agent planning for NEOs 1999 IMPACT Workshop 1