HTN-MAKER: Learning HTNs with Minimal Additional Knowledge - - PowerPoint PPT Presentation

▶
htn maker learning htns with minimal additional knowledge
SMART_READER_LITE
LIVE PREVIEW

HTN-MAKER: Learning HTNs with Minimal Additional Knowledge - - PowerPoint PPT Presentation

Jul 17, 2023 158 likes •773 views

HTN-MAKER: Learning HTNs with Minimal Additional Knowledge Engineering Required Chad Hogg 1 Hctor Muoz-Avila 1 Ugur Kuter 2 1 Department of Computer Science & Engineering Lehigh University Bethlehem, Pennsylvania 18015, USA 2 University

slide-1
SLIDE 1

HTN-MAKER: Learning HTNs with Minimal Additional Knowledge Engineering Required

Chad Hogg 1 Héctor Muñoz-Avila 1 Ugur Kuter 2

1Department of Computer Science & Engineering

Lehigh University Bethlehem, Pennsylvania 18015, USA

2University of Maryland

Institute for Advanced Computer Studies College Park, Maryland 20742, USA

23rd International AAAI Conference on Artificial Intelligence 2008-07-16

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-2
SLIDE 2

Motivation

Planning systems require some planning knowledge Simple STRIPS planners need only operators, which describe how actions change the world More advanced planners require operators and additional planning knowledge for problem solving Problem Developing additional planning knowledge for complex planners is a difficult manual process performed by domain experts

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-3
SLIDE 3

Motivation

Planning systems require some planning knowledge Simple STRIPS planners need only operators, which describe how actions change the world More advanced planners require operators and additional planning knowledge for problem solving Problem Developing additional planning knowledge for complex planners is a difficult manual process performed by domain experts

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-4
SLIDE 4

Motivation

Planning systems require some planning knowledge Simple STRIPS planners need only operators, which describe how actions change the world More advanced planners require operators and additional planning knowledge for problem solving Problem Developing additional planning knowledge for complex planners is a difficult manual process performed by domain experts

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-5
SLIDE 5

Motivation (2)

Objective We describe an algorithm for learning such planning knowledge for a certain class of advanced planners known as Hierarchical Task Network (HTN) planners

Actually, a specific form of HTN planning known as Ordered Task Decomposition (OTD) OTD planners are useful because of potential for efficiency and expressivity; used in most fielded applications

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-6
SLIDE 6

Motivation (2)

Objective We describe an algorithm for learning such planning knowledge for a certain class of advanced planners known as Hierarchical Task Network (HTN) planners

Actually, a specific form of HTN planning known as Ordered Task Decomposition (OTD) OTD planners are useful because of potential for efficiency and expressivity; used in most fielded applications

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-7
SLIDE 7

Motivation (2)

Objective We describe an algorithm for learning such planning knowledge for a certain class of advanced planners known as Hierarchical Task Network (HTN) planners

Actually, a specific form of HTN planning known as Ordered Task Decomposition (OTD) OTD planners are useful because of potential for efficiency and expressivity; used in most fielded applications

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-8
SLIDE 8

STRIPS Planning

States are sets of logical atoms describing the world Operators describe actions, when they may be undertaken, and how they affect the world The objective is to find a sequence of actions (a plan) that, when executed from a given initial state, will result in a state that satisfies certain given goals Example Operator :operator ((!load-truck ?pkg ?trk ?loc) :preconditions ((at ?pkg ?loc) (at ?trk ?loc)) :add-effects ((at ?pkg ?loc)) :del-effects ((in ?pkg ?trk)))

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-9
SLIDE 9

STRIPS Planning

States are sets of logical atoms describing the world Operators describe actions, when they may be undertaken, and how they affect the world The objective is to find a sequence of actions (a plan) that, when executed from a given initial state, will result in a state that satisfies certain given goals Example Operator :operator ((!load-truck ?pkg ?trk ?loc) :preconditions ((at ?pkg ?loc) (at ?trk ?loc)) :add-effects ((at ?pkg ?loc)) :del-effects ((in ?pkg ?trk)))

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-10
SLIDE 10

STRIPS Planning

States are sets of logical atoms describing the world Operators describe actions, when they may be undertaken, and how they affect the world The objective is to find a sequence of actions (a plan) that, when executed from a given initial state, will result in a state that satisfies certain given goals Example Operator :operator ((!load-truck ?pkg ?trk ?loc) :preconditions ((at ?pkg ?loc) (at ?trk ?loc)) :add-effects ((at ?pkg ?loc)) :del-effects ((in ?pkg ?trk)))

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-11
SLIDE 11

STRIPS Planning

States are sets of logical atoms describing the world Operators describe actions, when they may be undertaken, and how they affect the world The objective is to find a sequence of actions (a plan) that, when executed from a given initial state, will result in a state that satisfies certain given goals Example Operator :operator ((!load-truck ?pkg ?trk ?loc) :preconditions ((at ?pkg ?loc) (at ?trk ?loc)) :add-effects ((at ?pkg ?loc)) :del-effects ((in ?pkg ?trk)))

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-12
SLIDE 12

HTN Planning

Tasks are symbolic representations of activities that should be accomplished Methods describe how to decompose complex tasks into sequences of simpler

  • nes

The objective is to find a hierarchy of decompositions from a given sequence of tasks down to a sequence of actions that can be executed from the given initial state Example Method :method ((deliver ?pkg ?loc) :preconditions ((at ?pkg ?src) (at ?trk ?src) (same-city ?src ?loc)) :subtasks ((!load-truck ?pkg ?trk ?src) (deliver ?pkg ?loc)))

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-13
SLIDE 13

HTN Planning

Tasks are symbolic representations of activities that should be accomplished Methods describe how to decompose complex tasks into sequences of simpler

  • nes

The objective is to find a hierarchy of decompositions from a given sequence of tasks down to a sequence of actions that can be executed from the given initial state Example Method :method ((deliver ?pkg ?loc) :preconditions ((at ?pkg ?src) (at ?trk ?src) (same-city ?src ?loc)) :subtasks ((!load-truck ?pkg ?trk ?src) (deliver ?pkg ?loc)))

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-14
SLIDE 14

HTN Planning

Tasks are symbolic representations of activities that should be accomplished Methods describe how to decompose complex tasks into sequences of simpler

  • nes

The objective is to find a hierarchy of decompositions from a given sequence of tasks down to a sequence of actions that can be executed from the given initial state Example Method :method ((deliver ?pkg ?loc) :preconditions ((at ?pkg ?src) (at ?trk ?src) (same-city ?src ?loc)) :subtasks ((!load-truck ?pkg ?trk ?src) (deliver ?pkg ?loc)))

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-15
SLIDE 15

HTN Planning

Tasks are symbolic representations of activities that should be accomplished Methods describe how to decompose complex tasks into sequences of simpler

  • nes

The objective is to find a hierarchy of decompositions from a given sequence of tasks down to a sequence of actions that can be executed from the given initial state Example Method :method ((deliver ?pkg ?loc) :preconditions ((at ?pkg ?src) (at ?trk ?src) (same-city ?src ?loc)) :subtasks ((!load-truck ?pkg ?trk ?src) (deliver ?pkg ?loc)))

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-16
SLIDE 16

Learning HTN Methods from STRIPS Plans

Annotated Tasks We define the notion of an annotated task as a task with effects (what it means to accomplish the task) and preconditions (when it is possible to accomplish the task) Example: (deliver ?pkg ?loc) has effect (at ?pkg ?loc) Given A STRIPS domain D, a collection of annotated tasks T, a planning problem P in D, and a solution plan S to P, Find Method definitions for the tasks in T whose subtasks are actions from S or tasks from T

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-17
SLIDE 17

Learning HTN Methods from STRIPS Plans

Annotated Tasks We define the notion of an annotated task as a task with effects (what it means to accomplish the task) and preconditions (when it is possible to accomplish the task) Example: (deliver ?pkg ?loc) has effect (at ?pkg ?loc) Given A STRIPS domain D, a collection of annotated tasks T, a planning problem P in D, and a solution plan S to P, Find Method definitions for the tasks in T whose subtasks are actions from S or tasks from T

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-18
SLIDE 18

Learning HTN Methods from STRIPS Plans

Annotated Tasks We define the notion of an annotated task as a task with effects (what it means to accomplish the task) and preconditions (when it is possible to accomplish the task) Example: (deliver ?pkg ?loc) has effect (at ?pkg ?loc) Given A STRIPS domain D, a collection of annotated tasks T, a planning problem P in D, and a solution plan S to P, Find Method definitions for the tasks in T whose subtasks are actions from S or tasks from T

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-19
SLIDE 19

Main Steps of the HTN-MAKER Algorithm

1

Traverse the actions in the input plan, identifying states in which the effects of an annotated task have become satisfied

2

Regress annotated task effects backward through plan to each previous state where preconditions of annotated task are satisfied

3

Create a new method for the annotated task. Operators that contribute to the accomplishment of the effects become subtasks of this new method, as may the tasks associated with previously-learned methods

4

The preconditions of this new method are the union of task preconditions and preconditions of subtasks that are not provided by earlier subtasks

5

Learned methods are stored so that their tasks may be used as subtasks in methods that will be learned later

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-20
SLIDE 20

Main Steps of the HTN-MAKER Algorithm

1

Traverse the actions in the input plan, identifying states in which the effects of an annotated task have become satisfied

2

Regress annotated task effects backward through plan to each previous state where preconditions of annotated task are satisfied

3

Create a new method for the annotated task. Operators that contribute to the accomplishment of the effects become subtasks of this new method, as may the tasks associated with previously-learned methods

4

The preconditions of this new method are the union of task preconditions and preconditions of subtasks that are not provided by earlier subtasks

5

Learned methods are stored so that their tasks may be used as subtasks in methods that will be learned later

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-21
SLIDE 21

Main Steps of the HTN-MAKER Algorithm

1

Traverse the actions in the input plan, identifying states in which the effects of an annotated task have become satisfied

2

Regress annotated task effects backward through plan to each previous state where preconditions of annotated task are satisfied

3

Create a new method for the annotated task. Operators that contribute to the accomplishment of the effects become subtasks of this new method, as may the tasks associated with previously-learned methods

4

The preconditions of this new method are the union of task preconditions and preconditions of subtasks that are not provided by earlier subtasks

5

Learned methods are stored so that their tasks may be used as subtasks in methods that will be learned later

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-22
SLIDE 22

Main Steps of the HTN-MAKER Algorithm

1

Traverse the actions in the input plan, identifying states in which the effects of an annotated task have become satisfied

2

Regress annotated task effects backward through plan to each previous state where preconditions of annotated task are satisfied

3

Create a new method for the annotated task. Operators that contribute to the accomplishment of the effects become subtasks of this new method, as may the tasks associated with previously-learned methods

4

The preconditions of this new method are the union of task preconditions and preconditions of subtasks that are not provided by earlier subtasks

5

Learned methods are stored so that their tasks may be used as subtasks in methods that will be learned later

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-23
SLIDE 23

Main Steps of the HTN-MAKER Algorithm

1

Traverse the actions in the input plan, identifying states in which the effects of an annotated task have become satisfied

2

Regress annotated task effects backward through plan to each previous state where preconditions of annotated task are satisfied

3

Create a new method for the annotated task. Operators that contribute to the accomplishment of the effects become subtasks of this new method, as may the tasks associated with previously-learned methods

4

The preconditions of this new method are the union of task preconditions and preconditions of subtasks that are not provided by earlier subtasks

5

Learned methods are stored so that their tasks may be used as subtasks in methods that will be learned later

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-24
SLIDE 24

Example

Input is initial state (S0), plan, and annotated tasks

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-25
SLIDE 25

Example (2)

Use initial state and plan to fill in intermediate states

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-26
SLIDE 26

Example (3)

The effects of (deliver ?p ?l) become true in S3 with ?p = P , ?l = L2, so learn some methods starting with the previous Initial method has effects of annotated task as preconditions, no subtasks

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-27
SLIDE 27

Example (3)

The effects of (deliver ?p ?l) become true in S3 with ?p = P , ?l = L2, so learn some methods starting with the previous Initial method has effects of annotated task as preconditions, no subtasks

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-28
SLIDE 28

Example (4)

The (!unload P T L2) operator fulfills the outstanding precondition, so remove it, add the preconditions of the

  • perator to the method, and make it a subtask

This is now a method describing how to deliver a package from a certain set of states

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-29
SLIDE 29

Example (4)

The (!unload P T L2) operator fulfills the outstanding precondition, so remove it, add the preconditions of the

  • perator to the method, and make it a subtask

This is now a method describing how to deliver a package from a certain set of states

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-30
SLIDE 30

Example (5)

Now we go back one step further in the plan, starting again with an empty method

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-31
SLIDE 31

Example (6)

The method that we already learned fulfills the outstanding precondition (through the minimal effects of its annotated task), so remove it, add preconditions of method, and make it a subtask

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-32
SLIDE 32

Example (7)

The (!drive T L1 L2 C) operator fulfills an outstanding precondition, so remove it, add the preconditions of the

  • perator to the method, and make it a subtask

This is now a method describing how to deliver a package from an additional set of states

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-33
SLIDE 33

Example (7)

The (!drive T L1 L2 C) operator fulfills an outstanding precondition, so remove it, add the preconditions of the

  • perator to the method, and make it a subtask

This is now a method describing how to deliver a package from an additional set of states

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-34
SLIDE 34

Example (8)

Now we go back to the first step in the plan

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-35
SLIDE 35

Example (9)

The new previously learned method is useful in the same way

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-36
SLIDE 36

Example (10)

The (!load P T L1) operator fulfills a precondition, so we add it as well This is a third method, allowing us to accomplish the task from yet another set of states

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-37
SLIDE 37

Example (10)

The (!load P T L1) operator fulfills a precondition, so we add it as well This is a third method, allowing us to accomplish the task from yet another set of states

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-38
SLIDE 38

Theoretical Results

We say that a task t can be made equivalent to the goals G

  • f a classical planning problem by making G the annotated

effects of t Planning with methods learned by HTN-MAKER is sound, relative to the goals equivalent to the tasks There exists a finite number of input problems in a domain such that planning with methods learned by HTN-MAKER

  • n them is complete, relative to the annotated tasks

provided HTN Planning with a learned domain is more expressive than STRIPS planning

Classically-partitionable problems, which are sequences of (state, goal) pairs, are an example of a class of problems that can be solved using learned methods but cannot be expressed in classical planning

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-39
SLIDE 39

Theoretical Results

We say that a task t can be made equivalent to the goals G

  • f a classical planning problem by making G the annotated

effects of t Planning with methods learned by HTN-MAKER is sound, relative to the goals equivalent to the tasks There exists a finite number of input problems in a domain such that planning with methods learned by HTN-MAKER

  • n them is complete, relative to the annotated tasks

provided HTN Planning with a learned domain is more expressive than STRIPS planning

Classically-partitionable problems, which are sequences of (state, goal) pairs, are an example of a class of problems that can be solved using learned methods but cannot be expressed in classical planning

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-40
SLIDE 40

Theoretical Results

We say that a task t can be made equivalent to the goals G

  • f a classical planning problem by making G the annotated

effects of t Planning with methods learned by HTN-MAKER is sound, relative to the goals equivalent to the tasks There exists a finite number of input problems in a domain such that planning with methods learned by HTN-MAKER

  • n them is complete, relative to the annotated tasks

provided HTN Planning with a learned domain is more expressive than STRIPS planning

Classically-partitionable problems, which are sequences of (state, goal) pairs, are an example of a class of problems that can be solved using learned methods but cannot be expressed in classical planning

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-41
SLIDE 41

Theoretical Results

We say that a task t can be made equivalent to the goals G

  • f a classical planning problem by making G the annotated

effects of t Planning with methods learned by HTN-MAKER is sound, relative to the goals equivalent to the tasks There exists a finite number of input problems in a domain such that planning with methods learned by HTN-MAKER

  • n them is complete, relative to the annotated tasks

provided HTN Planning with a learned domain is more expressive than STRIPS planning

Classically-partitionable problems, which are sequences of (state, goal) pairs, are an example of a class of problems that can be solved using learned methods but cannot be expressed in classical planning

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-42
SLIDE 42

Theoretical Results

We say that a task t can be made equivalent to the goals G

  • f a classical planning problem by making G the annotated

effects of t Planning with methods learned by HTN-MAKER is sound, relative to the goals equivalent to the tasks There exists a finite number of input problems in a domain such that planning with methods learned by HTN-MAKER

  • n them is complete, relative to the annotated tasks

provided HTN Planning with a learned domain is more expressive than STRIPS planning

Classically-partitionable problems, which are sequences of (state, goal) pairs, are an example of a class of problems that can be solved using learned methods but cannot be expressed in classical planning

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-43
SLIDE 43

Evaluation

Experiments measuring coverage in Logistics-Transportation, Blocks-World, and Satellite domains

These are common benchmarks from planning competitions and the planning literature

Begin with HTN domain containing no methods, learn from solutions to a set of training problems, attempt to solve test problems with the learned methods Increase size of training set linearly from 1 to 75 problems Results averaged over 5 random problem distributions Logistics-Transportation and Satellite show rapid convergence to nearly complete domains, Blocks-World does not

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-44
SLIDE 44

Evaluation

Experiments measuring coverage in Logistics-Transportation, Blocks-World, and Satellite domains

These are common benchmarks from planning competitions and the planning literature

Begin with HTN domain containing no methods, learn from solutions to a set of training problems, attempt to solve test problems with the learned methods Increase size of training set linearly from 1 to 75 problems Results averaged over 5 random problem distributions Logistics-Transportation and Satellite show rapid convergence to nearly complete domains, Blocks-World does not

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-45
SLIDE 45

Evaluation

Experiments measuring coverage in Logistics-Transportation, Blocks-World, and Satellite domains

These are common benchmarks from planning competitions and the planning literature

Begin with HTN domain containing no methods, learn from solutions to a set of training problems, attempt to solve test problems with the learned methods Increase size of training set linearly from 1 to 75 problems Results averaged over 5 random problem distributions Logistics-Transportation and Satellite show rapid convergence to nearly complete domains, Blocks-World does not

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-46
SLIDE 46

Evaluation

Experiments measuring coverage in Logistics-Transportation, Blocks-World, and Satellite domains

These are common benchmarks from planning competitions and the planning literature

Begin with HTN domain containing no methods, learn from solutions to a set of training problems, attempt to solve test problems with the learned methods Increase size of training set linearly from 1 to 75 problems Results averaged over 5 random problem distributions Logistics-Transportation and Satellite show rapid convergence to nearly complete domains, Blocks-World does not

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-47
SLIDE 47

Evaluation

Experiments measuring coverage in Logistics-Transportation, Blocks-World, and Satellite domains

These are common benchmarks from planning competitions and the planning literature

Begin with HTN domain containing no methods, learn from solutions to a set of training problems, attempt to solve test problems with the learned methods Increase size of training set linearly from 1 to 75 problems Results averaged over 5 random problem distributions Logistics-Transportation and Satellite show rapid convergence to nearly complete domains, Blocks-World does not

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-48
SLIDE 48

Evaluation

Experiments measuring coverage in Logistics-Transportation, Blocks-World, and Satellite domains

These are common benchmarks from planning competitions and the planning literature

Begin with HTN domain containing no methods, learn from solutions to a set of training problems, attempt to solve test problems with the learned methods Increase size of training set linearly from 1 to 75 problems Results averaged over 5 random problem distributions Logistics-Transportation and Satellite show rapid convergence to nearly complete domains, Blocks-World does not

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-49
SLIDE 49

Results

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-50
SLIDE 50

Evaluation (2)

Since paper submission we have been working on planning efficiency results Currently only have results for Logistics-Transportation For small problems, FastForward is much faster than HTN planner with learned domain For large problems, advantages of HTN planning

  • vercome lack of focus in domain
  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-51
SLIDE 51

Evaluation (2)

Since paper submission we have been working on planning efficiency results Currently only have results for Logistics-Transportation For small problems, FastForward is much faster than HTN planner with learned domain For large problems, advantages of HTN planning

  • vercome lack of focus in domain
  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-52
SLIDE 52

Evaluation (2)

Since paper submission we have been working on planning efficiency results Currently only have results for Logistics-Transportation For small problems, FastForward is much faster than HTN planner with learned domain For large problems, advantages of HTN planning

  • vercome lack of focus in domain
  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-53
SLIDE 53

Evaluation (2)

Since paper submission we have been working on planning efficiency results Currently only have results for Logistics-Transportation For small problems, FastForward is much faster than HTN planner with learned domain For large problems, advantages of HTN planning

  • vercome lack of focus in domain
  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-54
SLIDE 54

Results (2)

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-55
SLIDE 55

Conclusions

We have described a new algorithm, HTN-MAKER, for learning HTN domain knowledge from classical planning problems and their solutions We have presented theoretical results showing that:

The methods learned by HTN-MAKER are sound and complete relative to the set of goals for which annotated tasks are provided The learned methods can be used to solve problems that could not be expressed using the classical planning knowledge from which they were learned

Our experiments in three well-known planning domains demonstrated that HTN-MAKER converged to a set of HTN methods that could solve nearly all problems in the domain

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-56
SLIDE 56

Conclusions

We have described a new algorithm, HTN-MAKER, for learning HTN domain knowledge from classical planning problems and their solutions We have presented theoretical results showing that:

The methods learned by HTN-MAKER are sound and complete relative to the set of goals for which annotated tasks are provided The learned methods can be used to solve problems that could not be expressed using the classical planning knowledge from which they were learned

Our experiments in three well-known planning domains demonstrated that HTN-MAKER converged to a set of HTN methods that could solve nearly all problems in the domain

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-57
SLIDE 57

Conclusions

We have described a new algorithm, HTN-MAKER, for learning HTN domain knowledge from classical planning problems and their solutions We have presented theoretical results showing that:

The methods learned by HTN-MAKER are sound and complete relative to the set of goals for which annotated tasks are provided The learned methods can be used to solve problems that could not be expressed using the classical planning knowledge from which they were learned

Our experiments in three well-known planning domains demonstrated that HTN-MAKER converged to a set of HTN methods that could solve nearly all problems in the domain

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-58
SLIDE 58

Conclusions

We have described a new algorithm, HTN-MAKER, for learning HTN domain knowledge from classical planning problems and their solutions We have presented theoretical results showing that:

The methods learned by HTN-MAKER are sound and complete relative to the set of goals for which annotated tasks are provided The learned methods can be used to solve problems that could not be expressed using the classical planning knowledge from which they were learned

Our experiments in three well-known planning domains demonstrated that HTN-MAKER converged to a set of HTN methods that could solve nearly all problems in the domain

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-59
SLIDE 59

Conclusions

We have described a new algorithm, HTN-MAKER, for learning HTN domain knowledge from classical planning problems and their solutions We have presented theoretical results showing that:

The methods learned by HTN-MAKER are sound and complete relative to the set of goals for which annotated tasks are provided The learned methods can be used to solve problems that could not be expressed using the classical planning knowledge from which they were learned

Our experiments in three well-known planning domains demonstrated that HTN-MAKER converged to a set of HTN methods that could solve nearly all problems in the domain

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER

slide-60
SLIDE 60

Conclusions (2)

Future Work Refine methods through merging or Reinforcement Learning-based selection Study effect of providing more or less information in annotated tasks Thank You Questions?

  • C. Hogg, H. Muñoz-Avila, U. Kuter

HTN-MAKER