CSC2542 Bernhard Nebel, and Jussi Rintanen. Introduction to - PDF document

Acknowledgements Some of the slides used in this course are modifications of Dana Nau’s lecture slides for the textbook Automated Planning, licensed under the Creative Commons Attribution-NonCommercial-ShareAlike License: http://creativecommons.org/licenses/by-nc-sa/2.0/ Other slides are modifications of slides developed by Malte Helmert, CSC2542 Bernhard Nebel, and Jussi Rintanen. Introduction to Planning I have also used some material prepared by P@trick Haslum. I would like to gratefully acknowledge the contributions of these researchers, Sheila McIlraith and thank them for generously permitting me to use aspects of their Department of Computer Science presentation material. University of Toronto Fall 2010 1 2 02 Clamp board 03 Establish datum point at bullseye (0.25, 1.00) 004 B VMC1 0.10 0.34 01 Install 0.15-diameter side-milling tool [a representation] of future behavior … 02 Rough side-mill pocket at (-0.25, 1.25) length 0.40, width 0.30, depth 0.50 usually a set of actions, with temporal and 03 Finish side-mill pocket at (-0.25, 1.25) other constraints on them, for execution by length 0.40, width 0.30, depth 0.50 04 Rough side-mill pocket at (-0.25, 3.00) some agent or agents. length 0.40, width 0.30, depth 0.50 plan n. - Austin Tate 05 Finish side-mill pocket at (-0.25, 3.00) 4. A drawing or diagram made to scale length 0.40, width 0.30, depth 0.50 [ MIT Encyclopedia of the Cognitive Sciences , 1999] showing the structure or arrangement 004 C VMC1 0.10 1.54 01 Install 0.08-diameter end-milling tool 1. A scheme, program, or method of something. [...] worked out beforehand for the 004 T VMC1 2.50 4.87 01 Total time on VMC1 accomplishment of an objective: a 5. In perspective rendering, one of 005 A EC1 0.00 32.29 01 Pre-clean board (scrub and wash) plan of attack. several imaginary planes 02 Dry board in oven at 85 deg. F perpendicular to the line of vision 2. A proposed or tentative project or 005 B EC1 30.00 0.48 01 Setup 02 Spread photoresist from 18000 RPM spinner between the viewer and the object course of action: had no plans for the 005 C EC1 30.00 2.00 01 Setup being depicted. evening. 02 Photolithography of photoresist using phototool in "real.iges" 6. A program or policy stipulating a 3. A systematic arrangement of elements 005 D EC1 30.00 20.00 01 Setup service or benefit: a pension plan. or important parts; a configuration or 02 Etching of copper 005 T EC1 90.00 54.77 01 Total time on EC1 outline: a seating plan; the plan of a Synonyms: blueprint, design, project, story. scheme, strategy 006 A MC1 30.00 4.57 01 Setup 02 Prepare board for soldering 006 B MC1 30.00 0.29 01 Setup A portion of a manufacturing process plan 3 4 02 Screenprint solder stop on board 006 C MC1 30 00 7 50 01 Setup

Modes of Planning Example Planning Applications � Mixed Initiative Planning Automated Plan Generation � 5 6 Autonomous Agents for Space Exploration Other Autonomous Systems � Autonomous planning, scheduling, control � NASA: JPL and Ames � Remote Agent Experiment (RAX) � Deep Space 1 � Mars Exploration Rover (MER) Not necessarily embodied! 7 8

Manufacturing Automation Other Autonomous Systems: � Sheet-metal bending machines - Amada Corporation � Software to plan the sequence of bends [Gupta and Bourne, J. Manufacturing Sci. and Engr. , 1999] Not necessarily embodied! 9 9 10 Other Applications Games E.g., Bridge Baron - Great Game Products � Scheduling with Action Choices & Resource Requirements � 1997 world champion of computer bridge � Problems in supply chain management [Smith, Nau, and Throop, AI Magazine , 1998] � HSTS (Hubble Space Telescope scheduler) � 2004: 2nd place Us:East declarer, West dummy � Workflow management Opponents:defenders, South & North Finesse(P 1 ; S) � Air Traffic Control Contract:East – 3NT East: ♠ KJ74 On lead:West at trick 3 � Route aircraft between runways and terminals. Crafts West: ♠ A2 Out: ♠ QT98653 must be kept safely separated. Safe distance depends LeadLow(P 1 ; S) FinesseTwo(P 2 ; S) on craft and mode of transport. Minimize taxi and wait time. PlayCard(P 1 ; S, R 1 ) EasyFinesse(P 2 ; S) StandardFinesse(P 2 ; S) BustedFinesse(P 2 ; S) � Character Animation West — ♠ 2 … … � Generate step-by-step character behaviour from high- (North — ♠ Q) (North — � 3) level spec StandardFinesseTwo(P 2 ; S) StandardFinesseThree(P 3 ; S) FinesseFour(P 4 ; S) � Plan-based Interfaces � E.g. NLP to database interfaces PlayCard(P 2 ; S, R 2 ) PlayCard(P 3 ; S, R 3 ) PlayCard(P 4 ; S, R 4 ) PlayCard(P 4 ; S, R 4 ’ ) � Plan recognition North — ♠ 3 East — ♠ J South — ♠ 5 South — ♠ Q 11 12

Other Applications (cont.) Outline � Web Service Composition � Conceptual model for planning � Compose web services, and monitor their execution � Classes of planning problems � Many of the web standards have a lot of connections to � Classes of planners and example instances action representation languages � Beyond planning � BPEL; BPEL-4WS allow workflow specifications � Planning research – the big picture � DAML-S allows process specifications � Business Process Composition /Workflow Management � Some of what I hope you’ll get from the course � Including Grid Services/Scientific Workflow Management � Genome Rearrangement � The relationship between different organisms can be measured by the number of “evolution events” (rearrangements) that separate their genomes � Find shortest (or most likely) sequence of rearrangements between a pair of genomes 13 14 State Transition System Conceptual Model s 1 s 0 Σ = ( S,A,E , γ ) 1. Environment put � S = {states} take � A = {actions} location 1 location 2 location 1 location 2 move2 move1 move2 move1 � E = {exogenous events} s 3 s 2 � State-transition function put γ : S x ( A ∪ E ) → 2 S take � S = {s 0 , …, s 5 } State transition system location 1 location 2 location 1 location 2 System Σ � A = { move1, move2, Σ = ( S,A,E , γ ) load unload put, take, load, unload } s 4 s 5 S = {states} � E = {} move2 A = {actions} � γ : see the arrows E = {exogenous events} move1 γ = state-transition function Dock Worker Robots (DWR): location 1 location 2 location 1 location 2 15 16

Conceptual Model Conceptual Model 2. Controller 3. Planner’s Input Planning problem Planning problem Planning problem Planner Given observation Controller Omit unless o in O , produces Observation function planning is online action a in A h : S → O s 3 location 1 location 2 17 18 s 1 s 0 put Conceptual Model Planning Problem take 4. Planner’s Output P = ( Σ , s 0 ,G ) location 1 location 2 location 1 location 2 move1 move1 move2 move2 Σ : System Description s 3 s 2 put Planner Instructions to S 0 : Initial state(s) the controller take E.g., Initial state = s 0 location 1 location 2 location 1 location 2 unload load G: Objective s 4 s 5 Goal state, move2 Set of goal states, Set of tasks, move1 “trajectory” of states, location 1 location 2 location 1 location 2 Objective function, … E.g., Goal state = s 5 The Dock Worker Robots (DWR) domain 19 20

s 1 s 0 Outline Plans put take Classical plan : take � Conceptual model for planning location 1 location 2 location 1 location 2 a sequence of actions � Classes of planning problems move1 move1 move1 move2 move2 E.g., 〈 take, move1, load, move2 〉 � Classes of planners and example instances s 3 s 2 put � Beyond planning Policy : � Planning research – the big picture partial function from S into A take � Some of what I hope you’ll get from the course E.g., location 1 location 2 location 1 location 2 load load {(s 0 , take), unload ( s 1 , move1), s 4 s 5 ( s 3 , load), move2 move2 (s 4 , move2)} move1 location 1 location 2 location 1 location 2 The Dock Worker Robots (DWR) domain 21 22 Different Classes Planning Problems Different Classes Planning Problems dynamics: deterministic , nondeterministic, probabilistic Varying components of the planning problem specification yields different classes of problems. E.g., observability: full, partial, none dynamics: deterministic, nondeterministic, probabilistic horizon: finite , infinite objective requirement: satisfying , optimizing observability: full, partial, none … horizon: finite, infinite � classical planning objective requirement: satisfying, optimizing � conditional planning with full observability … � conditional planning with partial observability � conformant planning � markov decision processes (MDP) � partial observable MDP (POMDP) � preference-based/over-subscription planning 23 24