Ehsan Nazerfard

1
Learning Search Control Rules for Plan-Space
Planner
Yong QuSuresh KatukamSubbarao Kambhampati

• Presented by
• Ehsan Nazerfard

2
Outline

• Related works
• Overview of UCPOP
• EBL in UCPOP
• Schematic of UCPOPEBL
• Failures and Explanations
• Regression and Propagation
• Example
• Generalization
• Performance

3
Related Works

• SNLPEBL
• Suresh Katukam and Subbarao Kambhampati
• UCPOPEBL
• Yong Qu and Subbarao Kambhampati
• GraphPlanEBL
• ? and Subbarao Kambhampati

4
Overview of UCPOP

• Flowchart of refinement process in UCPOP

5
Overview of UCPOP

• The Briefcase Domain
• Classic Problem
• Goal
• Getting an empty briefcase to the office while
  leaving everything else at home
• Initial state
• Paycheck P is in briefcase and briefcase is at
  home

6
Trace of UCPOP Solving the Problem
7
Explanation Based Learning

• explaining why something is a good idea
  (and generalizing from that)
• construct an explanation
• generalize the explanation (introduce variables
  into the explanation)
• construct a new rule from the generalized
• Utility Problem

8
EBL in UCPOP

• Learn control rules from search failures
• Analytical failure
• Cross depth limit failure
• To avoid failures in similar situations in future
• Online learning
• Offline learning
• Control rules
• In Selection form
• In Rejection form

9
Schematic of UCPOPEBL

• The Explanation is a minimal set of
  constraints in P that are together
  inconsistent
• Regression and Propagation are used to abstract
  failure info to higher level plans in tree
• DFS is usually adopted

10
Failures and Explanations

• Some are detected by consistency checks
• (s1lts2 , s2lts1) ? O)
• (p_at_s1 , p_at_s1) ? A)
• Implicit failures may only be detected by use of
  domain specific axioms
• An object can not be at two places at the same
  time
• Domain axiom in briefcase domain
• no block can have another block on it and be
  clear
• Domain axiom in blocks world domain

11
Regression and Propagation

• Initial failure explanation E is regressed up
  through decision d leading to failing
  plan
• Compute condition E' that need to be true in plan
  before d such that failure will result after d
• Its possible to create rule that rejects d
  whenever E' is true but ...

12
Regression and Propagation

• For regression its useful to think UCPOP
  decisions as STRIPS operators

• Regressing constraints over decisions

13
Regression and Propagation

• When all the branches under P are failing, we can
  construct an explanation for P itself
• Sometimes its not needed to wait for all
  branches to fail
• This is called dependency directed backtracking
  (ddb)

14
Trace of UCPOPEBL Solving the Problem
15
UCPOPEBL solving the problem

• Explanation of failure for p7 that is regressed
  over step-addition, take-out (p)

• E6 E5 E4 E3 E2 are constructed in such manner
• Explanation of failure for p3

16
UCPOPEBL solving the problem

• By regressing E3 over establishment under p2

• this leads to useful control rule
• establishment of closed(B)_at_G should be avoided
  when the paycheck is in the
  office, briefcase is not at office and we want
  the briefcase to be at office and paycheck to
  be left at home

17
Generalization

• replace any problem-specific constants with
  variables without affecting its correctness
• Only those binding that are forced by initial
  explanation of the failure.
• Step names
• Object names
• Final search control rule

18
Rule Storage

• Generalized rules are available to the planner
  via rule corpus
• Learning phase
• Subsequent planning episodes
• Some checks is done by the planner on rule
  corpus

19
Performance

• Performance of UCPOPEBL in Blocks world and
  Briefcase Domain
• 100 random problems were generated for each
  domain

20
Other issues

• Learn useful rules from Domain axioms

• Expressive action representation can obviate the
  need to specific domain knowledge to
  some extent.
• for example

21
Conclusion

• extend control rule learning framework to UCPOP
• (previous work on SNLPEBL)
• Explanation, Regression, Propagation and
• rule constructing
• Solving classic problem from Briefcase Domain
• Reduce the need for domain specific failure
  theories using expressive action
  representation

22
Any Question would be welcome
Thanks for your attention