GIPO II: HTN Planning in a Toolsupported Knowledge Engineering Environment

1
GIPO II HTN Planning in a Tool-supported
Knowledge Engineering Environment

• Lee McCluskey
• Donghong Liu
• Ron Simpson
• Department of Computing and
• Mathematical Sciences,
• The University of Huddersfield

2
Summary

• In this paper/talk, we postulate a domain
  independent method supported by
• static/dynamic tools
• dual structuring of the model via object
  hierarchies and HTN operators
• to acquire, develop and validate a hierarchical
  planning domain model.

3
Contents

• Contents
• Background Development Method
• Modelling
• The Transparency Tool
• The HyHTN planner / algorithm
• Conclusions

4
1. Background Development Method

• Our main interest is in Knowledge Engineering -
  the acquisition, validation and modelling of AI
  Planning knowledge.
• We are creating experimental research platforms
  for investigating
• the integration of planning tools
• methods of planning knowledge acquisition

5
Advertisements

• GIPO-I and GIPO-II software can be obtained
  freely for Linux, Solaris and Windows via our
  website
• http//scom.hud.ac.uk/planform/gipo/
• There is a demonstration of GIPO II at the Demo
  session later on this evening
• Also please note that there is a NEW
  comprehensive web site for planners and
  schedulers, planning tools, domain models - on
• http//scom.hud.ac.uk/planet/repository/
• sponsored by the EU PLANET Network

6
Recap on GIPO (related papers in ECP01,
ISMIS02,AIPS02)

• An open tools environment supporting domain
  acquisition and domain modelling integrated via a
  GUI around the Object Centred language OCLh and
  associated method.
• Our approach has been to stick to a classical
  planning foundation and concentrate on the
  integration/acquisition/modelling aspects

7
GIPO - I

• GIPO version 1 contains within a GUI
• domain model editors
• static analysis checking tools
• a plan stepper
• an exporter to PDDL
• planners API for 3rd party planners
• a solution animator
• a random task generator
• (new in 02) an induction tool to help in
  operator acquisition

8
Domain Model Development with GIPO-II Method
Outline

• GIPO-II is designed for writing HTN models. GIPO
  II contains
• domain model editors
• basic static analysis checking tools
• is object class hierarchy consistent?
• do object state descriptions satisfy invariants?
• are predicate structures and operator schema
  mutually consistent?
• are task specifications consistent with the
  domain model?
• a plan stepper
• a solution visualiser
• PLUS
• Transparency Tool HyHTN Planner
• TO ADD
• Induction and Pattern tools

Acquisition of Objects/ Object State
Behaviour -use GIPO-II GUI or (next release) use
generic paterns
Static Analysis Tools
Operator Acquisition Using GIPO-II GUI or (next
release) induction
Static Analysis including Transparency Tool
Solving simple tasks using The Plan Stepper
Solving more complex Tasks with HyHTN
Solution visualier
9
2. Modelling

• GIPOs modelling based on the idea of engineering
  a planning domain so that the universe of
  potential states of objects are defined first
• objects grouped within classes under a class
  hierarchy.
• Each class in the hierarchy may have a
  behaviour'' in the sense that objects of that
  class have changeable properties and relations.
• An object may inherit behaviour from each class
  above it in the hierarchy.
• This proceeds before operator definition and
  makes it possible to accurately induce operator
  schema (as shown in our AIPS02 paper..)

GIPO Example from a TransLog Domain
10
Object hierarchy example
physical_obj
at
package
vehicle
waiting certified delivered
moveable available
railv
attached unattached
train
traincar
11
Operator representation

• The basis of both primitive and hierarchical
  operators are transitions of objects LHS gt RHS
• Depending on how precisely specified the LHS/RHS
  of transitions are, this abstraction uniformly
  encompasses goal conditions, pre-conditions,
  necessary and conditional effects, deterministic
  and non-deterministic actions..

12
Method Representation

• Primitive operators are composed of transitions
  and constraints. But hierarchical domain models
  contain methods - HTN-like operators
• Methods are defined using
• transitions
• constraints
• a task network (of methods and achieve goals)

13
Task Representation

• Tasks are defined using
• an initial state
• a task network (of methods and achieve goals)
• constraints on variables/orders in the task
  network
• EXAMPLE (without the initial state data)
• ( achieve(ss(traincar,traincar1,
• at(traincar1,city1-ts1))),
• transport(pk-5-z,city3-cl1-z,city2-cl1),
• achieve(ss(package,pk-5,
• at(pk-5,X),delivered(pk-5) )) ,
• before(1,3),
• serves(X,city3-x) )

14
(No Transcript)
15
3. The Transparency Tool

• A Methods transitions can be viewed as (or
  translated to) pre- and post conditions for that
  methods task network
• For Example

ACHIEVE GOAL
task network
POST
PRE
POST
PRE
POST
POST
PRE
task network CAN BE EXPANDED BY EXPANDING EACH
METHOD
POST
PRE
16
The Transparency Property

• Methods are regulated by the semantic property of
  transparency -- this ensures they are structured
  in a coherent manner.
• A method is sound if its task network necessarily
  achieves its POST-condition with respect to the
  objects in the methods transitions.
• The transparency property is then as follows
• A method m is transparent if it and every
  expansion of m, consistent with its static
  constraints, is sound.

17
The Transparency Tool

• The tool is executed after each method has been
  created or updated.
• An error shows that the methods task network
  cannot perform the methods transitions.
• Its use can uncover subtle errors in methods ..
  But it cannot spot all errors!

18
(No Transcript)
19
4. The HyHTN planner / algorithm

• A methods task networks can be composed of all
  achieve goals, all methods, or somewhere in
  between. The idea is to use HyHTN at any point in
  this development space.
• It inputs an OCLh domain model and task as shown
  above.
• It outputs a plan that is input to the plan
  visualiser, displaying the decompositions making
  up the solution to the task.

20
HyHTN Algorithm

• HYHTN is a forward state advancing HTN planner.
  This has the advantage that heuristic state-space
  search can be used to establish achieve-goal'
  conditions.
• Thus the performance of eg SHOP-like algorithms
  in HTN planning, and the performance of fast
  forward algorithms in pre-condition planning have
  been combined into a flexible, efficient hybrid
  system.
• Also the transparency property reduces the
  possibility of choosing methods that lead to dead
  - ends, as every task network decomposition that
  satisfies its static constraints is guaranteed to
  achieve its post-conditions.

21
Experiments with HyHTN

• In the paper we include a comparison of HyHTN vs
  SHOP using the Translog Domain with object
  classes such as cities, regions, packages,
  trucks, trains, planes, cranes, ramps etc.
• The domain model contains 34 parameterised
  methods and 58 parameterised primitive operator
  structures. The SHOP model is of a similar size.
• The specific problems concern the transport of up
  to 10 packages, with 5 connected cities, 15
  locations, 15 cranes to maintain one crane at
  each location, and 11 trucks in one location in
  the initial state.The packages were of different
  types bulky, liquid, granular, and mail.

22
(No Transcript)
23
5. Summary

• We have introduced a method which includes two
  powerful tools for HTN planning and domain
  development
• Transparency to ensure that hierarchical
  operators do what their transitions say they do.
  This is static validation.
• HyHTN a planner that can be used to help
  develop hierarchical operators in that initially
  it can be run with achieve goals and these can
  be replaced by canned plans as the domain model
  is refined.

24
6. Future

• GIPO-II is still only usable by experts we want
  to make it more accessible -
• We intend to incorporate the transparency
  property into an induction and theory revision
  tool we are creating for inducing HTN models from
  examples.
• We intend to transfer GIPO into a Web Service and
  eventually make it into an autonomous knowledge
  acquisition agent for planning problems.