CBR for Design

1
CBR for Design

• Upmanyu Misra
• CSE 495

2
Design Research

• Develop tools to aid human designers
• Automate design tasks
• Better understanding of design
• Increase quality
• Take lesser time
• Improve predictability of design

REUSE
3
Reuse of Design

• Reuse old design share intellectual property
  (IP)
• As the reuse increases, the complexity
  increases
• Human assistance is mandatory
• Directed towards assisting human designer rather
  than making intelligent decisions by own

4
Design Task

• Routine design
• is completely a part of a set of potential
  designs
• all variables, their ranges, and knowledge to
  compute their values are directly derivable from
  the set
• easily implemented
• Innovative Design
• is partially derived from the set of potential
  designs
• all components need to be derived. The knowledge
  is incomplete
• design needs to be iteratively derived
• Creative Design
• no overlap with the set of potential design. The
  set needs to be extended
• All components need to be defined

5
Design Task (figure)
6
Approaches for design tasks
Routine Only

• Formulae
• Constraints
• Rules and grammars
• CBR
• Prototype based reasoning

- Goel (1989), Domeshek and Kolodner (1992),
Hinrichs (1992)
7
The PCM Model

• Propose involves using domain knowledge to map
  part or all of the specification to partial or
  complete design proposals
• Critique assessment of the proposed design
  solution
• Modify takes info about a failure of a proposed
  design as its input and then changes the design
  to get closer to the desired specification

8
The PCM Model
The CBR Cycle
9
Mapping Design Task to CBR-cycle
10
Case Based Design

• Defined as the process of creating a new design
  solution by combining and/or adapting previous
  design solution(s)
• useful tool for intelligent system design in a
  domain where either an explicit model does not
  exist or one is not yet adequately understood
• can learn from interaction with user

11
A Framework for CBD Systems
12
Characteristics of CBD System

• Can produce complete and complex designs based on
  relatively small knowledge base
• design starts from complete cases, implicitly
  achieving trade-offs between several constraints
• design history of existing cases makes design
  problem solving more efficient
• using cases as a source of knowledge allows
  learning by storing new cases

13
CBR System Architecture

• Four Knowledge Containers
• Vocabulary should be able to capture all salient
  features of the design. Task dependent
• Case base
• - usage cases can capture both regular/normal
  situations as well as exceptions/abnormal
  situations
• - granularity for task-oriented user support,
  the grain size of the cases matches that of the
  decisions made
• Similarity measures to compare queries and cases
  in their corresponding representations
• Solution transformation contains knowledge
  required to evaluate solutions

14
Case Retrieval for Innovative/Creative design

• Flexible case retrieval
• Structural similarity assessment
• Similarity assessment in terms of adaptability

15
Case Retrieval

• Flexible Case Retrieval Given a large case
  base, a problem, and a number of aspects that are
  relevant for similarity assessment, a set of
  cases is to be retrieved which show similar
  aspects as in the actual problem
• The aim is to exploit different views on single
  cases
• Importance of certain aspects for similarity
  assessment may not be known at memory
  organization time
• - dynamic weighting is required
• - use kd trees, Case Retrieval Nets etc.

16
Fish Shrink Algorithm

• Used for Flexible Case Retrieval
• Selects and ranks potential cases from a large
  set of cases
• Considers different aspects (representations) of
  cases
• Main idea it should be more efficient to avoid
  searching in the nearby neighborhood of cases
  which have already been found to be inappropriate

17
Fish Shrink

• A representation function takes the
  case and outputs the aspect in the desired
  representation space
• case 1 (20, empty, 0.05)
• case 2 (19, half-empty, 0.9)
• A distance function that can take two
  representations in space and calculate
  the distance of the two cases in this aspect

18
Fish Shrink
19
Fish Shrink Method

• View distance SD
• The view distance from the query to some case is
  called test distance, and the view distance
  between two cases is called the base distance
• This is a basic distance function, researchers
  generally use their own
• Presumption View distance function have to
  satisfy the triangle equality

20
Fish Shrink Algorithm
21
Fish and Shrink
0
T3
T1
T2
1
Distance to the query
22
Structural Similarity Assessment and Adaptation
Using Graphs

• To retrieve structurally most similar cases
• Structured case representation ? Graph
• Find maximum common sub graph
• CAD example for industrial building
• Object represented by set of attributes
  describing its geometry and type
• Different pipe system shows different topological
  relations
• Building structure can be mapped onto its pipe
  system

23
Required Functionality

• A compile function is used to translate the
  selected attributes and their relations into
  graphs
• A recompile function is used to translate the
  selected solution graphs into their
  attributes-based representation
• Retrieving case is conducted by selecting the
  case having maximum common sub graph with the
  problem
• Structural adaptation proceeds by combining case
  parts that are not included in the sub graph

24
Structural similarity assessment and adaptation

• A graph g(V,E), where V is the set of vertex and
• mcs(G) is the maximum common sub graph of a set
  of graphs G
• Let be the set of all graphs, O be the a
  finite set of objects represented by attribute
  values and other relationships, P( ) be the
  power set of
• compile
• recompile
• retrieve
• match
• adapt

25
Set of
match
adapt
mcs
cases
retrieve
set of solution_g
problem_g
case base_g
recompile
compile
compile
problem_a
case base_a
set of solution_a
26
Example of structural similarity assessment TOPO

• Consider geometric neighborhoods as well as
  structural similarity
• Compile and Recompile
• There can be several kinds of relations for
  different types
• Retrieval
• Use Fish and Shrink algorithm
• Search for maximum clique instead of maximum
  common sub graph
• Search clique in one graph representing all
  possible matches between two graphs, combination
  graph
• Adaptation
• Sub graphs that are not in the clique are added
  to the solution under user defined constraint

27
Combination Graph

• Nodes in the combination graph is the matching of
  relationships in original graph
• Two nodes are connected together if the two
  matching does not contradict each other
• clique- finding is done by Bron and Kerboshs
  algorithm, by extending complete sub graph of
  size k to k1 by adding vertices connected to all
  vertices in the found sub graph

28
Graph f
Graph g
b1
b3
R1(b,a)
R5(b,b)
R6(b,a)
R10(b,b)
a2
a1
a4
a3
R4(a,a)
R9(a,b)
R2(a,b)
R3(a,b)
R7(b,a)
R8(a,b)
b2
b4
Combination graph
result
R4(a,a)ltgt R9(a,a)
b3
R6(b,a)
R5(b,b)ltgt R10(b,b)
R10(b,b)
R1(b,a)ltgt R7(b,a)
b1
a4
a3
R1(b,a)
R5(b,b)
R9(a,b)
a2
a1
R7(b,a)
R8(a,b)
R4(a,a)
b4
R2(a,b)ltgt R8(a,b)
R3(a,b)
R1(b,a)ltgt R6(b,a)
R2(a,b)
b2
R3(a,b)ltgt R8(a,b)
29
Structural Adaptation by Case Combination

• Example EADOC, supporting conceptual design
  phase in designing aircraft panel structure
• User specifies initial requirements, objectives,
  and preferences
• Specific plans for certain model is not available
• Partial model for evaluating behavior are
  available
• Four CBR cycles, each results in a set of
  solution that can serve as the input for next
  cycle
• Additional information is needed to guide the
  retrieval
• Solutions can be biased to previous tasks

30
EADOCS Design Process
specifications
prototype
selection
initial target
prototype
target
remaining
retrieve part

precedent case
concept
retrieve case
selection
case base
yes
no
concept
structural
remaining
adaptation
target
concept
precedent case
modification
concept
optimization
31
Summary
32
References

• JÄorg Walter Schaaf, Fish Shrink. A next step
  towards efficient case retrieval in large scaled
  case bases, EWCBR96
• Ian Watson, Srinath Parera, Case-Based Design A
  Review and Analysis of Building Design
  Applications, Journal of AI for Engineering
  Design, Analysis and Manufacturing, 1997
• Katy Borner, CBR for Design, ???