Practical HTN Planning Putting HTN Planning into Use Literature Human Planning � Klein, G. (1998) Sources of Power: How People Make Decisions, MIT Press. � Refinement Search � Kambhampati, S., Knoblock, C.A. and Yang, Q. (1995) Planning as Refinement Search: A Unified � Framework for Evaluating Design Tradeoffs in Partial-Order Planning, Artificial Intelligence, Vol. 76, No. 1- 2, pp. 167-238, Elsevier. Nonlin � http://www.aiai.ed.ac.uk/project/nonlin/ � Tate, A. (1977) Generating Project Networks, Proceedings of the Fifth International Joint Conference on � Artificial Intelligence (IJCAI-77) pp. 888-893, Boston, Mass. USA, August 1977. O-Plan � http://www.aiai.ed.ac.uk/project/oplan/ � Currie, K and Tate, A. (1991) O-Plan: the Open Planning Architecture, Artificial Intelligence Vol. 52, No. 1, � pp 49-86, Elsevier. Other Practical Planners � Ghallab, M., Nau, D. and Traverso, P., Automated Planning – Theory and Practice , chapters 19, 22 and � 23. Elsevier/Morgan Kaufmann, 2004. Practical HTN Planning 2 1
Overview � Human Approaches to Planning � Practical HTN Planning � Refinement Planning as a Unifying View � Nonlin and O-Plan Features � QA (Modal Truth Criterion) � Time, Resource and Other Constraint Handling � I-X/I-Plan Overview Practical HTN Planning 3 Some Planning Features Expansion of a high level abstract plan into greater detail � where necessary. High level ‘chunks’ of procedural knowledge (Standard � Operating Procedures, Best Practice Processes, Tactics Techniques and Procedures, etc.) at a human scale - typically 5-8 actions - can be manipulated within the system. Ability to establish that a feasible plan exists, perhaps for a � range of assumptions about the situation, while retaining a high level overview. Analysis of potential interactions as plans are expanded or � developed. Identification of problems, flaws and issues with the plan. � Deliberative establishment of a space of alternative options, � perhaps based on different assumptions about the situation involved, of especial use ahead of time, in training and rehearsal, and to those unfamiliar with the situation or utilising novel equipment. Practical HTN Planning 4 2
More Planning Features Monitoring of the execution of events as they are expected to � happen within the plan, watching for deviations that indicate a necessity to re-plan (often ahead of this becoming a serious problem). Represent the dynamic state of the world at points in the plan � and use this for ‘mental simulation’ of the execution of the plan. Pruning of choices according to given requirements or � constraints. Situation dependent option filtering (sometime reducing the � choices normally open to one ‘obvious’ one. Satisficing search to find the first suitable plan that meets the � essential criteria. Heuristic evaluation and prioritisation of multiple possible � choices within the constrained search space. Uniform use of a common plan representation with � embedded rationale to improve plan quality, shared understanding, etc. Practical HTN Planning 5 Human Approach � Previous slides describe aspects of problem solving behaviour observed in expert humans working in unusual or crisis situations. � Gary Klein, “Sources of Power”, MIT Press, 1998. � But they also describe the hierarchical and mixed initiative approach to planning in AI developed over the last 30 years. Practical HTN Planning 6 3
HTN - Planning Approach � HTN Planning is a useful paradigm… � Compose workflows/processes from requirements and component/template libraries � Covers simple through to very complex (pre- planned) components � Allows for execution support, reactive repair, recovery, etc. � Suited to mixed initiative (people and systems) planning and execution � Gives an understandable framework within which specialised constraint solvers, domain- specific planners (e.g. route finders), optimisers, plan analysers and simulators can work Practical HTN Planning 7 HTN - Activity Composition Plan Library A2 Refinement S2 S1 “Initial” Plan “Final” Plan A2.1 A2.2 A2 Refine A1 A4 A5 A1 A4 A5 A3 A3 Introduce activities to achieve preconditions Resolve interactions between conditions and effects Handle constraints (e.g. world state, resource, spatial, etc.) Practical HTN Planning 8 4
HTN – Initial Plan as “Goals” Plan Library Ax Refinement S1 S2 P “Initial” Plan “Refined” Plan P P Refine A1.1 A1.2 Q Q Initial Plan can be any combination of Activities and Constraints Practical HTN Planning 9 Nonlin (1974-1977) � Hierarchical Task Network Planning � Partial Order Planner � Plan Space Planner � Goal structure-based plan development - considers alternative “approaches” only based on plan rationale � QA/Modal Truth Criterion Condition Achievement � Condition “Types” to limit search � “Compute Conditions” for links to external data and systems (attached procedures) � Time and Resource Constraint checks � Nonlin core is basis for text book descriptions of HTN Planning Practical HTN Planning 10 5
Nonlin Domain Language – TF opschema makeon actschema puton pattern {on $*x $*y} pattern {put $*x on top of $*y} expansion 1 goal {cleartop $*x} conditions usewhen {cleartop $*x} at self 2 goal {cleartop $*y} usewhen {cleartop $*y} at self 3 action {put $*x on top of $*y} usewhen {on $*x $*z} at self orderings 1 ---> 3 2 ---> 3 effects + {on $*x $*y} vars x undef y undef; - {cleartop $*y} “typed” condition restricts search space end; - {on $*x $*z} example of search control knowledge + {cleartop $*z} opschema makeclear vars x undef y undef z undef; pattern {cleartop $*x} end; expansion 1 goal {cleartop $*y} 2 action {put $*y on top of $*z} always {cleartop table}; orderings 1 ---> 2 conditions usewhen {on $*y $*x} at 2 initially {on c a} usewhen {cleartop $*z} at 2 {on a table} vars x <:non table:> y undef {on b table} z <:et <:non $*x:> <:non $*y:> :>; {cleartop c} end; {cleartop b} ; plan goal {on a b} goal {on b c}; $*x is a variable Practical HTN Planning 11 QA/Modal Truth Criterion � QA in a partially ordered network of nodes � Way to establish value of a condition P=V at some point in the plan � Yes/no/maybe responses � Alternative Terminology: • Contributors, deletors (Austin Tate, Nonlin, QA, Edinburgh, 1975-7) • White nights and clobberers (David Chapman, MIT, MTC, 1987, 1 st Formalisation) • Producers, consumers (Some textbooks) � Initially just allowed imposition of orderings on nodes for a condition, a � b (ordering) � Later also allowed variables within condition to be constrained – = (codesignation), ≠ (non-codesignation) � Intuitively, a white knight is an activity which re-establishes a clobbered precondition p � A clobberer in a plan can be "defeated" by imposing ordering or codesignation/non-codesignation constraints on the plan, or by inserting a white knight between the clobberer and the point where a condition is needed Practical HTN Planning 12 6
QA/Modal Truth Criterion Before After Need to ensure no deletor appears between a chosen contributor and point of need Contributor No Effect Deletor P=V Practical HTN Planning 13 O-Plan (1983-1999) Features Hierarchical Task Network Planning � Nonlin-like goal-structure, QA and Typed/Compute conditions � Partial-Plan “Refinement “ Approach � Plan State has “flaws”/issues attached � Agenda Architecture with Plan Modification Operations � “Opportunistic Search” (agenda type, branch1/branch N) � Multiple constraint managers with yes/no [and maybe] results � Least Commitment Approach (on activity ordering, � object/variable bindings and other constraints) Constraint “Posting” rather than explicit commitments � (and/or trees with sets of “before” temporal constraints and variable binding (= and ≠ ) constraints) [as in MOLGEN] Goal structure recording and monitoring to preserve plan � rationale Practical HTN Planning 14 7
O-Plan (1983-1999) Features Practical HTN Planning 15 O-Plan Domain Language – TF types objects = (a b c table), movable_objects = (a b c); always {cleartop table}; schema puton; vars ?x = ?{type movable_objects}, ?y = ?{type objects}, ?z = ?{type objects}; vars_relations ?x /= ?y, ?y /= ?z, ?x /= ?z; expands {puton ?x ?y}; only_use_for_effects {on ?x ?y} = true, “typed” condition restricts search space {cleartop ?y} = false, example of search control knowledge {on ?x ?z} = false, {cleartop ?z} = true; conditions only_use_for_query {on ?x ?z} achieve {cleartop ?x} achieve {cleartop ?y}; end_schema; ?x is a variable Practical HTN Planning 16 8
O-Plan Agent Architecture Practical HTN Planning 17 O-Plan Agent Architecture Later became Plan Modification Later became Operators • Issues • Nodes • Constraints • Annotations Practical HTN Planning 18 9
Recommend
More recommend
