Different Combination of Multiple Pattern Database

1
Different Combination of Multiple Pattern Database
Presented by Yunping Wang CMPUT 652 Project
2
Outline

• Problem Definition
• Background
• Approach and Experiment Setting
• Preliminary Results
• Future Plan
• Reference

3
Problem Definition

• Motivation
• How best to use a fixed amount of memory for
  storing pattern database?
• Which performance is better?
• Using n pattern database of total size is m
• OR
• Using one pattern database of size m

4
Background

• IDA
• Richard Korf introduced Iterative Deepening A
  (IDA) in 1985.
• The idea is to trade space for time use less
  space (a lot less space!) but have the program
  run a bit slower.

5
Background

• IDA
• for a Node N, define
• g(N)the distance from the start state to N.
• h(N)an at most estimate of the distance from N
  to the goal state
• f(N)g(N)h(N)

6
Background

• Pattern Database
• PDB a heuristic stored as lookup table
• Created by abstracting the state space
• Guaranteed to be a lower bound
• Information stored in PDB
• abstracted states
• h value( the states distance to goal)

7
Approach

• How to get H value from multiple pattern
  database?
• It is admissible if every h is admissible.

8
Experiment Setting

• Total memory 10,080
• PDB selecting from

9
Experiment Setting

• PDB selection
• Do not select duplicate PDB
• Test PDB should be complementing

10
Experiment Setting

• Puzzle Type
• 8-Puzzle from class
• Testing states
• 200 start states (randomly selected) for each
  combination

11
Experiment Setting

• Avoiding Unnecessary Lookup
• IDA depth bound7
• g(n) 3
• Stop doing PDB lookups as soon as hgt4 is
  found.
• Might result in extra IDA iterations
• h1(n) 5 next bound is 8
• h2(n) 7 next bound is 10

12
Preliminary Results --nodes generated

• One big PDB generating the biggest nodes
• Multiple smaller PDBs helps reduce the nodes
  generated

13
Preliminary Results --CPU time

• No clear trend with different PDB combinations.
• CPU time is determined by the PDB combinations
  and the swap cost between PDBs.

14
Preliminary Results -- nodes
generated in different group
15
Future Plan

• The optimal strategy to partition the PDB.
• Combination
• --- try different combination
• Size
• --- set the total memory in other size number

16
Reference

• J. Culberson and J. Schaeffer. Searching with
  Pattern Databases, CSCSI 96 (Canadian AI
  Conference), Advances in Artificial Intelligence,
  Springer-Verlag, pp.402416, 1996.
• R. Korf. Finding Optimal Solutions to Rubiks
  Cube Using Pattern Databases, Proceedings of the
  Fourteenth National Conference on Artificial
  Intelligence (AAAI-97), pp. 700705, 1997.
• R. Holte, Newton. Multiple Pattern Database,
  Unpublished manuscript, 2001.

17
Thank for Your Attention! Question?