1 usc cs 541 ai planning lecture notes yolanda gil plan representation and reasoning with...
TRANSCRIPT
1USC CS 541 AI Planning Lecture Notes Yolanda Gil
Plan Representation and Reasoningwith Description Logics
and Ontologies
Yolanda Gil
Lecture Notes, October 4, 2000
CS 541 Artificial Intelligence Planning
www.isi.edu/~gil/cs541
2USC CS 541 AI Planning Lecture Notes Yolanda Gil
Outline
Representing actions and plans with description logic Action taxonomies (CLASP) Plan taxonomies (SUDO-PLANNER) Goal taxonomies (EXPECT)
Planning ontologies Process Specification Language (PSL), NIST PLANET
3USC CS 541 AI Planning Lecture Notes Yolanda Gil
Representing Knowledge in Description Logic (DL)
Description logics are extensions of frame-based systems where classes can be defined intensionally
Ex: SUVs are vehicles with 4 seats that weight between 1T and 2T
Class taxonomy is automatically generated through subsumption A subsumes B iff all instances of B are also instances of A
Instances can be automatically classified Ex: MyNewCar is a vehicle with 4 seats that weighs 1.3T => MyNewCar is an SUV
Relations can also have definitions and can be classified Tradeoff between expressivity and efficient reasoning Some well know description logic systems: CLASSIC,
LOOM, NIKL
4USC CS 541 AI Planning Lecture Notes Yolanda Gil
Representing Planning Knowledge in Description Logics: Overview
Action taxonomies in CLASP extended language to represent action networks
Plan taxonomies in SUDO-PLANNER plan subsumption of partially ordered plans
Goal taxonomies in EXPECT expressive representations of goals and their
parameters
These systems can exploit the descriptions of all the objects in the domain (domain knowledge) in order to reason about action, goal, and plan descriptions
5USC CS 541 AI Planning Lecture Notes Yolanda Gil
CLASP: CLAssification of Scenarios and Plans [Devanbu and Litman 94] Extension of a DL system (CLASSIC)
Language to express action networks– Sequence, loop, repeat, test, subplan
Subsumption and classification algorithms for that language– Action network subsumption viewed as DFA acceptance
Propositional, STRIPS-style representation of actions States (goals are represented as states) Actions Plans Scenarios (plan instances)
Reasoning based on these descriptions: Organizing plan classes Retrieving plan types Validation of scenarios
6USC CS 541 AI Planning Lecture Notes Yolanda Gil
Core Classes in CLASP
(DEFINE-CONCEPT Action (PRIMITIVE (AND Classic-Thing
(AT-LEAST 1 Actor) (ALL ACTOR Agent) (EXACTLY 1 PRECONDITION) (ALL PRECONDITION State) (EXACTLY 1 ADD-LIST) (ALL ADD-LIST State) (EXACTLY 1 DELETE-LIST) (ALL DELETE-LIST State)
(EXACTLY 1 GOAL) (ALL GOAL STATE))))
(DEFINE-CONCEPT State (PRIMITIVE Classic-Thing)) (DEFINE-PLAN Plan (PRIMITIVE
(AND Clasp-Thing (EXACTLY 1 INITIAL) (ALL INITIAL State)
(EXACTLY 1 GOAL) (ALL GOAL State) (EXACTLY 1 PLAN-EXPRESSION) (ALL PLAN-EXPRESSION (LOOP Action)))))
7USC CS 541 AI Planning Lecture Notes Yolanda Gil
Defining Actions, States and Plans in CLASP in a Telephony Domain
(DEFINE-CONCEPT System-Act (AND Action
(ALL ACTOR System-Agent)))
(DEFINE-CONCEPT Connect-Dialtone-Act (AND System-Act
(ALL PRECONDITION (AND Off-Hook-State Idle-State)) (All Add-LIST Dialtone-State) (ALL DELETE-LIST Idle-State (ALL GOAL (AND Off-Hook-State
Dialtone-State))))
(DEFINE-CONCEPT Callee-Off-Hook-State (PRIMITIVE State))(DEFINE-CONCEPT Callee-On-Hook-State (PRIMITIVE State))(DEFINE-CONCEPT Callee-Off-Caller-On-State (AND Callee-Off-Hook-State
Caller-On-Hook-State))
(DEFINE-PLAN Pots-Plan (AND Plan (ALL PLAN-EXPRESSION
(SEQUENCE (SUBPLAN Originate-And-Dial-Plan) (TEST (Callee-On-Hook-State
(SUBPLAN Terminate-Plan))(Callee-Off-Hook-State (SEQUENCE Non-Terminate-Act Caller-On-Hook-Act Disconnect Act)))))))
(DEFINE-PLAN Originate-And-Dial-Plan (AND Plan
(ALL PLAN-EXPRESSION(SEQUENCE Caller-Off-Hook-Act Connect-Dialtone-Act Dial-Digits-Act))))
8USC CS 541 AI Planning Lecture Notes Yolanda Gil
Defining Instances in CLASP
(CREATE SCENARIO pots-busy-scenario (AND Plan
(FILLS INITIAL state-u1on-u2off) (FILLS GOAL state-u1on)
(FILLS PLAN-EXPRESSION (caller-off-hook-u1 connect-dialtone-on-u1 dial-digits-u1-to-u2 non-terminate-on-u2 caller-on-hook-u1 disconnect-u1))))
(CREATE-IND state-u1on-u2off (AND state-U1on State-U2off)) (CREATE-IND connect-dialtone-on-u1 (AND Connect-Dialtone-Act (FILLS ACTOR switching-system) (FILLS PRECONDITION state-u1off-idle)))
9USC CS 541 AI Planning Lecture Notes Yolanda Gil
SUDO-PLANNER [Wellman 88]
Exploits subsumption to control the search during plan generation
Actions represented in DL (NIKL), organized in taxonomy
Plans represented as partially ordered sets of actions Eliminate search nodes whose plan is subsumed (dominated)
by other nodes SUDO-PLANNER had other features not discussed here:
Uncertainty reasoning and partial goal satisfaction Policy constraints that relate actions to external events Conditional effects Qualitative probabilistic networks
10USC CS 541 AI Planning Lecture Notes Yolanda Gil
Action Taxonomy in SUDO-PLANNER
(defconcept surgery :is (:and action (:the route invasive-path-into-body)))
(defconcept biopsy :is-primitive action ...))
(defconcept open-lung-biopsy :is (:and biopsy (:the route open-lung-path)))
(defconcept open-lung-path :is (:and invasive-path-into-body ...))
System deduces that open-lung-biopsy is a surgery
11USC CS 541 AI Planning Lecture Notes Yolanda Gil
Plan Representation and Subsumption in SUDO-PLANNER Plan is described as a set of action types
associated with identifiers [(surgery, id1) (CABG, id2)]
Plan is simplified if action subsumption and same id [(surgery, id1) (CABG, id1)] -> [(surgery, id1)]
Plan subsumption Action network viewed as bipartite graph matching
a1 a2 a5
a3 a4 a6
a1 a4 a5
a2 a3 a6
12USC CS 541 AI Planning Lecture Notes Yolanda Gil
A* a1 A*
a2 A*
a2 b7 A*
A* a1 b5 A*
a1 b5 A*
...
XA* = {ai…aj} ai subsumes aj when i<j
Eliminating Redundant Paths in Plan Space Search Dominance-based planning:
Generate new nodes by adding constraints to search nodes Derive dominance (i.e., subsumption) based on domain
knowledge Eliminate nodes in the plan graph that are dominated by others
13USC CS 541 AI Planning Lecture Notes Yolanda Gil
Reasoning about Goals in EXPECT [Swartout et al 98]
Highly declarative representation of goals Goals as verb-based expressions Rich language of goal parameter types
– Qualification parameters that describe the type of task– Intentional and extensional sets
Given a goal, matcher looks for methods (operators) that have the capability of achieving that goal can match variabilized goals can decompose goal into subgoals through reformulations
Goal representations have been used in several contexts: representing planning goals problem solving agent matchmaking
14USC CS 541 AI Planning Lecture Notes Yolanda Gil
Representing Goals in EXPECT
Represented as a case grammar (verb + roles) ex: ESTIMATE OBJ duration OF trip
Roles can be filled by: a specific instance: add OBJ 3 TO 5 a concept: compute OBJ (spec-of factorial) OF 7 a type of instance: divide OBJ number BY 2 extensional sets: find OBJ (spec-of maximum) OF (54 15
256) intensional sets: add OBJ (set-of number) find OBJ (set-of (spec-of violated-
constraint)) IN configuration
Roles filled by concepts express task qualification parameters declaratively
(compute-factorial ?n) -> (compute (obj (spec-of factorial)) (of number)))
15USC CS 541 AI Planning Lecture Notes Yolanda Gil
Matching Goals in EXPECT Desired goals and available capabilities are automatically translated to LOOM concepts Classifier is used to find most specific method capability that subsumes the posted goal
Self-organizing method taxonomy
movecargo
aircraft
OBJ
WITH
movecargo
truck
OBJ
WITH
movecargo
vehicle
OBJ
WITH
movecargo
C-140
OBJ
WITH
Goal: (move (OBJ (inst-of cargo)) (WITH C-140))
Method capability: (move (OBJ (inst-of cargo)) (WITH (inst-of aircraft)))
16USC CS 541 AI Planning Lecture Notes Yolanda Gil
Flexible Matching through Goal Reformulation
When no capability matches a posted goal, but more specific versions of the goal match ex: no method to estimate round-trip time (rtt) of a vehicle,
but there are methods to estimate rtt of aircraft and trucks Use descriptive knowledge to reformulate goal
reexpress goal into subgoals by breaking down one of the arguments
recombine the results of solving subgoals Conjunctive (disjunctive) subgoals produce conjunctive
(disjunctive) reformulations Types of reformulations
Covering reformulation: subgoals cover partitions of a class Set reformulation: subgoals iterate over elements of a set Input reformulation: subgoals handle each of the subtypes
17USC CS 541 AI Planning Lecture Notes Yolanda Gil
Find route from location1 to location2
Find egress route from Ryad to Kuwait city
Calculate RTTfor transport aircraft
Calculate round-trip time (RTT)for aircraft
Calculate RTTfor combat aircraft
A) Subsumption-based match: the posted goal is subsumed by a capability
B) Reformulation-based match: the posted goal can besatisfied by combining two or more existing capabilities
Find route from city1 to city2
Find route from location1 to location2
C) Reverse subsumption-based match: a capability can satisfy some aspect of the goal
Find addresses of US citizens in Kuwait
Find phone numbers of US citizens in Kuwait
D) Partial match: a capability is similar/related to the posted goal
Goal Matching in EXPECT
18USC CS 541 AI Planning Lecture Notes Yolanda Gil
Overview of Planning Ontologies
Why planning ontologies knowledge reuse knowledge sharing knowledge modeling
Process descriptions in PSL temporal constraints resources
Describing plans in PLANET can represent state-based, plan-based search,
hierarchical plans captures plan representations understandable by people
19USC CS 541 AI Planning Lecture Notes Yolanda Gil
Process Specification Language (PSL) [NIST, 99]
National Institute of Standards and Technology (NIST), Manufacturing Systems Division Academic and industrial collaborators
Proposed to Int’l Standards Organization (ISO) PSL core represents widely accepted
commitments activity, activity-occurrence, object, timepoint
PSL extensions accommodate possible shareable agreements
Contains axioms defining terms and constraints Available at http://www.mel.nist.gov/psl/
22USC CS 541 AI Planning Lecture Notes Yolanda Gil
PSL Core
Activity Generic activity: occurrences, interruptions,
nondeterministic, subactivities Ordering: ordering over activities, complex ordering
relations, junctions Objects
Resources: capacities, homogeneous sets, inventories, divisibility, usage, resource paths, pools, requirements, resource roles, substitutability
States: defined, constraints Timepoints
Duration theory, activity durations, temporal orderings
23USC CS 541 AI Planning Lecture Notes Yolanda Gil
PLANET: a PLAN Semantic nET [Gil & Blythe 00]
Capture unifying views on planning algorithms constraints, commitments, task templates, alternative choices state-based and objective-based goals operator-based and HTN-based plans
Represent manually created plans typically include unintended flaws (incomplete, unjustified,
inconsistent) Capture planning context
initial constraints (guide, user advice, preferences) and restraints initial state, constraints and goals may be incompatible
Available from http://www.isi.edu/expect/projects/planet/
24USC CS 541 AI Planning Lecture Notes Yolanda Gil
Some Terms Defined in PLANET
Planning problems Planning problem context: world state, desired goals,
external constraints Planning problem: candidate plans (rejected, feasible,
selected) Goals and effects
Goals: state-based goals, objective-based goals Human readable descriptions
Actions, operators, and tasks Plan task descriptions: plan task templates, plan tasks, Capabilities, preconditions, effects, subtasks, primitive tasks,
plan steps Plans
Commitments, sub-plans, planning level
25USC CS 541 AI Planning Lecture Notes Yolanda Gil
PLANET: An Ontology for Representing Plans
justified
complete
consistent
feasible
plan-commitments
plan-refinements
sub-plans
capabilityeffectspreconditions
planning-level
human-readable description
sub-taskstask-of
Plan-task-templatePlan-task-templatetask-template
commitments
accomplishesordering
temporal
planning-problems
initial-stateworld-state
desired-goals
external constraints
candidate-plans
rejected
feasible
selected
unexplored
planning - level
state-based-goal-spec
objective-based-goal-spec
resource-neededamount
when-needed
Plan-task-descriptionPlan-task-description
Plan-taskPlan-task
PlanPlan
Planning-problem-contextPlanning-problem-context
Goal-specificationGoal-specification
Resource-requirements
Resource-requirements