8th. June 2016Michaja Pressmar

Master Thesis 16.11.2015 – 16.5.2016 Examiner: Prof. Dr. Malte Helmert Supervisor: Dr. Martin Wehrle

Analysing and Combining Static Pruning Techniques for Classical Planning Tasks

1. Introduction

Overview

2

2. Pruning Techniques3. Synergies4. Experiments5. Conclusion

1. Introduction

3

● Variables

driveA B → driveB A → loadA unloadA loadB unloadB

A A

B B

1. Introduction

A B

4

● Operators

● Initial state

● Goal state

Planning Task

Plan:

A T B TA A B BB A A BA T B TA AB A B B

1. Introduction

A B

5

loadA driveA->B unloadB

1. Introduction

A BC

A A AB A A

C A A

C T AB T A

A T A

C A TA A T

B A T

C B AA B A

B B AA B TB B T

C B T

C B CA B C

B B C

B A BA A B

C A BA T BB T B

C T B

B B BA B B

C B B

A C AB C A

C C A

C C TA C T

B C T

C C BB C B

A C B

C C C

A C C B C C

B A CC A C

A A C

C T CA T C

B T C

C C BB C B

A C B

B T TA T T

6

State Explosion Problem

State Explosion Problem

Dynamic Pruning

Static Pruning

● Variables

● Operators

● Initial state

● Goal state

A A

B B

unloadB ...

1. Introduction 7

2. Pruning Techniques

Safe Abstraction

8

Redundant Operator Reduction Dominance Pruning

2. Pruning Techniques 9

Safe Abstraction

Original version by Helmert (2006)

Variant by Haslum (2007)

General idea:1. Abstract safe variables

2. Solve abstract task

3. Refine plan

2. Pruning Techniques 10

Safe Abstraction

● Variables

driveA B → driveB A → loadA unloadA loadB unloadB

A

B B

● Operators

● Initial state

● Goal state

Planning Task

A A

B

Abstract Planning Task

TA B

2. Pruning Techniques

Plan: loadA unloadB

11

Safe Abstraction

A B

driveA B→

TA BA T B TA A B B

Abstract Planning Task Planning TaskOriginal

2. Pruning Techniques 12

Redundant Operator Reduction

Researched by Haslum and Jonsson (2000)

General idea: More operators than necessary

Remove redundant operators

2. Pruning Techniques 13

Redundant Operator Reduction

A B

C

2. Pruning Techniques 14

Dominance Pruning

Optimality-preserving

Torralba and Hoffmann (2015)

Compute state simulation

Kissmann and Torralba (2015)

Identify subsumed transitions

Remove „bad“ operators

2. Pruning Techniques 15

Dominance Pruning

A T B TA A B BB A A B

driveA B → driveB A → loadA unloadA loadB ● Operators unloadB

16

3. Synergies

17

Synergy Effects

P P' P''

Pruning

Method 1

Pruning

Method II

3. Synergies

Which synergies exist? → Proven theorems

positive Synergy

negative

Synergy

18

Synergy Effects

SafeAbstraction

RedundantOperators

DominancePruning

3. Synergies

4. Experiments

Implementation in Fast Downward

19

IPC Benchmarks

20

Roadmap Experiments

1. Techniques separately

2. Combinations and Synergies

4. Experiments

Configurations

Best performance

Synergies separately

Synergies compared

4. Experiments 21

Safe Abstraction

Compared 2 Variants:

1886

1413+80

50%22%

Original Version

Coverage:

35% 58%Variant (Haslum)

Instances abstractable

Safe Variables per instance

Expanded States Planning Time

baseline (Greedy + FF)

Safe

Abs

trac

tion

4. Experiments 22

Redundant Operator Reduction

Coverage-24

1886

1413

Expanded States Planning Time

baseline (Greedy + FF)

Red

unda

nt O

pera

tor

Red

uctio

n

No Coverage Increase

4. Experiments 23

Dominance Pruning

Powerful

up to 50 – 80% operators pruned

, but comparatively slow

Incremental Computation & Early-Abort

Expanded States Planning Time

baseline (A* + LMCut)

Dom

inan

ce P

runi

ng

4. Experiments 24

Synergies separately

Single technique vs combination

All synergies occur on IPC domains

ROR vs SA+ROR

Pruned Operators

ROR

ROR

(afte

r SA

)

4. Experiments 25

Synergies compared

ROR+SA vs SA+ROR

Combination vs Combination

No coverage increase

Operators left Planning Time

ROR + SA

SA +

RO

R

5. Conclusion

26

less strict Safe Abstraction more fine-grained ROR satisficing Dominance Pruning

Static Pruning can improve coverage Synergies exist & occur on IPC domains

Further research:

Thank you!