Post on 13-Apr-2017
Automating the Conceptual Design Process
Automated Design LaboratoryThe University of Texas at Austin
Mechanical Engineering Department
April 30, 2008Presented by Justin McKay
Automating the Systematic Design Process
Research Effort 1: Function Structure Grammar
Research Effort I1: Configuration Grammar
Research Effort III : Tree Search & Component Selection
Functional Specifications
Customer Needs
Black Box Model
Functional Decomposition
Function Structure
Configuration Design
Configuration Flow Graph
Component Selection
Analysis and
Evaluation
* Final * Design
prior work – automated concept generation
Function Structure
Functional Synthesis by Rule Execution
Design Grammar
Rules
Configuration Flow Graph’sFunction Structure
Functional Synthesis by Rule Execution
Configuration Flow Graph’s
• Graph Grammars – Function to Form mapping
Function Structure – describes the functions necessary to translate input flows into output flows
Graph Grammar - Rules for manipulating nodes and arcs within graph
Configuration Flow Graph – synthesizes individual or sets of components into a set of conceptual design configurations
Deriving Grammar RulesThe Sharper Image
Hybrid “Shake” 7-LED Magnetic-Induction Flashlight
The Bill of Materials
Housing Assembly
Cap Seals (O-rings)
Cover
KnobHous
ing
LED Assembly
Cover
Lenses
HousingLEDs
Circuit Board
Screws
Power Source Assembly
CoilCap
Cushion 1
Wire 1
Cushion 2Magnet
Wire 2
Wire 3
Switch
Circuit Board
Capacitor(Reverse Side)
Function Structure (the Seed)
Arc=flow type
Node=Function
Function to Form Mapping via Grammar Rules
Configuration Flow Graph
Function Structure
Grammar Rules
Derived Grammar Rules
LHS:Function Structure
RHS:Configuration Flow
GraphI
II
III
If LHS, then RHS
TK’s designer preference modeler (dpm)
1. observe designer’s decision making2. construct a model of designer preferences3. find best designs
Evaluate Designs
Generate Designs
Computer-to-Human Generation Feedback(heuristic sampling)
Human-to-Computer Evaluation Feedback
(designer preferences)
DPM
Designer SynthesisSoftware
Evaluate Designs
Generate Designs
Computer-to-Human Generation Feedback(heuristic sampling)
Human-to-Computer Evaluation Feedback
(designer preferences)
DPM
Designer SynthesisSoftware
Automated Concept Generation
feasible design space
candidate 1
candidate 2feasible design space
candidate 1
candidate 2result of generation
• multiple candidates• grammar generates 594 “bread slicer” designs• need to score and rank generated candidates
• i.e. Pugh’s Chart
DPM (designer preference modeler)
evaluation scheme
Generate CFG’s
Sample from solution space
Get designer’srating for selected concepts
Project concept ratings on to rule choices using recipes
feasible design spacefeasible design spacefeasible design spacefeasible design space
Compute “preference score”for each rule choice
A B C D
Score A Score C Score D Score B
Propagate “preference scores”over each recipe to compute overall worth of each design
Recipe A:
Rule #Rule # Rule #
. .
.
Recipe B:
Rule #Rule # Rule #
. .
.
Recipe C:
Rule #Rule # Rule #
. .
.
Recipe D:
Rule #Rule # Rule #
. .
.
Rule #1
Rule #2
Rule #3 . . .
Rule Score2
Rule Score1
Rule Score3
Rule Preference Scores:
feasible design spacefeasible design space
Figure 4: Illustration of design evaluation.
DPM (designer preference modeler)
Generate CFG’s
Sample from solution space
Get designer’srating for selected concepts
Project concept ratings on to rule choices using recipes
feasible design spacefeasible design spacefeasible design spacefeasible design space
Compute “preference score”for each rule choice
A B C D
Score A Score C Score D Score B
Propagate “preference scores”over each recipe to compute overall worth of each design
Recipe A:
Rule #Rule # Rule #
. .
.
Recipe B:
Rule #Rule # Rule #
. .
.
Recipe C:
Rule #Rule # Rule #
. .
.
Recipe D:
Rule #Rule # Rule #
. .
.
Rule #1
Rule #2
Rule #3 . . .
Rule Score2
Rule Score1
Rule Score3
Rule Preference Scores:
feasible design spacefeasible design space
Figure 4: Illustration of design evaluation.
Sampling Goals:• select a set to be
presented to the designer• reduce the number of
designer evaluations• capture the variety in
the design solution space
Generate CFG’s
Sample from solution space
Get designer’srating for selected concepts
Project concept ratings on to rule choices using recipes
feasible design spacefeasible design spacefeasible design spacefeasible design space
Compute “preference score”for each rule choice
A B C D
Score A Score C Score D Score B
Propagate “preference scores”over each recipe to compute overall worth of each design
Recipe A:
Rule #Rule # Rule #
. .
.
Recipe B:
Rule #Rule # Rule #
. .
.
Recipe C:
Rule #Rule # Rule #
. .
.
Recipe D:
Rule #Rule # Rule #
. .
.
Rule #1
Rule #2
Rule #3 . . .
Rule Score2
Rule Score1
Rule Score3
Rule Preference Scores:
feasible design spacefeasible design space
Figure 4: Illustration of design evaluation.
Seed of the design generation tree
31
3 13 10
13 10 22 31 22 29
29 35 29 35 31 35C1 C2 C3
C4 C5 C6 C7 C8 C9
Candidate
C1 (1,3,13)
C2 (3,13,22)
C3 (3,13,31)
C4 (1,3,10,29)
C5 (1,3,10,35)
C6 (3,10,22,29)
C7 (3,10,22,35)
C8 (3,10,29,31)
C9 (3,10,31,35)
Recipe
Recipe Pool (1,3,10,13,22,29,31,35)
Seed of the design generation tree
31
3 13 10
13 10 22 31 22 29
29 35 29 35 31 35C1 C2 C3
C4 C5 C6 C7 C8 C9
Seed of the design generation tree
31
3 13 10
13 10 22 31 22 29
29 35 29 35 31 35C1 C2 C3
C4 C5 C6 C7 C8 C9
Candidate
C1 (1,3,13)
C2 (3,13,22)
C3 (3,13,31)
C4 (1,3,10,29)
C5 (1,3,10,35)
C6 (3,10,22,29)
C7 (3,10,22,35)
C8 (3,10,29,31)
C9 (3,10,31,35)
Recipe
Recipe Pool (1,3,10,13,22,29,31,35)
1) compute the recipe pool
DPM (designer preference modeler)
2) calculate commonality measure of each candidate
3) select candidate with most commonly occurring rules and add it to the feedback set,
4) update pool, repeat steps 1-3
Frequency of Appearance of Each Rule in the Recipe Pool
3
9
6
3 3 3 3 330 0
3 3 30 0
30 0
30 0 0 00 0 0 0 0 0 0 0
1 3 10 13 22 29 31 35
Rule Number
Freq
uenc
y
First Iteration
Second Iteration
Third Iteration
Sampling Finished
Commonality Measure of Each Candidate
15 15 15
21 21 21 21 21 21
6 6
3
6 6
9
3 3 0
6
3 3 3 3 0 0 0 00 0 0 0 0 0 0 0 0
C1 C2 C3 C4 C5 C6 C7 C8 C9
Candidate
Com
mon
ality
Mea
sure
First Iteration
Second Iteration
Third Iteration
Sampling Finished
Generate CFG’s
Sample from solution space
Get designer ’srating for selected concepts
Project concept ratings on to rule choices using recipes
Compute “ preference score ”for each rule choice
Propagate “ preference scores ”over each recipe to compute overall worth of each design
DPM (designer preference modeler)
4) pair-wise comparison matrix for DFS
DPM (designer preference modeler)
Generate CFG’s
Sample from solution space
Get designer’srating for selected concepts
Project concept ratings on to rule choices using recipes
feasible design spacefeasible design spacefeasible design spacefeasible design space
Compute “preference score”for each rule choice
A B C D
Score A Score C Score D Score B
Propagate “preference scores”over each recipe to compute overall worth of each design
Recipe A:
Rule #Rule # Rule #
. .
.
Recipe B:
Rule #Rule # Rule #
. .
.
Recipe C:
Rule #Rule # Rule #
. .
.
Recipe D:
Rule #Rule # Rule #
. .
.
Rule #1
Rule #2
Rule #3 . . .
Rule Score2
Rule Score1
Rule Score3
Rule Preference Scores:
feasible design spacefeasible design space
Figure 4: Illustration of design evaluation.
Generate CFG’s
Sample from solution space
Get designer’srating for selected concepts
Project concept ratings on to rule choices using recipes
feasible design spacefeasible design spacefeasible design spacefeasible design space
Compute “preference score”for each rule choice
A B C D
Score A Score C Score D Score B
Propagate “preference scores”over each recipe to compute overall worth of each design
Recipe A:
Rule #Rule # Rule #
. .
.
Recipe B:
Rule #Rule # Rule #
. .
.
Recipe C:
Rule #Rule # Rule #
. .
.
Recipe D:
Rule #Rule # Rule #
. .
.
Rule #1
Rule #2
Rule #3 . . .
Rule Score2
Rule Score1
Rule Score3
Rule Preference Scores:
feasible design spacefeasible design space
Figure 4: Illustration of design evaluation.
6) extract how much the designer prefers a particular rule (i.e. function-to-form mapping) over others
DPM (designer preference modeler)
Generate CFG’s
Sample from solution space
Get designer’srating for selected concepts
Project concept ratings on to rule choices using recipes
feasible design spacefeasible design spacefeasible design spacefeasible design space
Compute “preference score”for each rule choice
A B C D
Score A Score C Score D Score B
Propagate “preference scores”over each recipe to compute overall worth of each design
Recipe A:
Rule #Rule # Rule #
. .
.
Recipe B:
Rule #Rule # Rule #
. .
.
Recipe C:
Rule #Rule # Rule #
. .
.
Recipe D:
Rule #Rule # Rule #
. .
.
Rule #1
Rule #2
Rule #3 . . .
Rule Score2
Rule Score1
Rule Score3
Rule Preference Scores:
feasible design spacefeasible design space
Figure 4: Illustration of design evaluation.
DPM (designer preference modeler)
7) calculate best/worst designs in the population
Design ProblemExample Problem Bread SlicerEvaluation Criteria Quality of Concepts
Design Generation and Sampling# of candidates generated 594# of rules in the recipe pool 30
# of candidates in DFS 12Design Evaluation
Recipe Pool1,7,8,10,11,12,13,15,17,20,21,22,23,24,25,26,
28,29,30,31,33,34,35,36,38,39,40,41,42,43
RPS-3,3,14,3,0,6,-3,7,-21,-7,15,3,24,-9,-14,7,
-3,3,-9,0,2,-9,0,3,0,0,9,9,9,-9 Recipe ScoreBest Candidate C155 10,11,15,22,23,29,38,39,40,41,42 67
Worst Candidate C498 1,8,11,13,17,38,39,43 -50
• 12 designs are selected as
feedback set
• finds best “bread slicer” in a population of 594
C155
DPM – results for bread slicer design
DPM – Summary and ImplicationsTK’s DPM Method1. Find minimum set of solutions (select candidates from candPool with high commonality measure)2. Of these, target candidates that best represents variety in the solution space
Limitations• Had to keep total # of rules small• Pair-wise comparison requires too many evaluations (66 evaluations for a DFS of 12 candidates)
Future Work• multi-criteria evaluation• learning from past “preference models” • use as “guiding” strategy when full space cannot be enumerated
Design ProblemExample Problem Bread SlicerEvaluation Criteria Quality of Concepts
Design Generation and Sampling# of candidates generated 594# of rules in the recipe pool 30
# of candidates in DFS 12Design Evaluation
Recipe Pool1,7,8,10,11,12,13,15,17,20,21,22,23,24,25,26,
28,29,30,31,33,34,35,36,38,39,40,41,42,43
RPS-3,3,14,3,0,6,-3,7,-21,-7,15,3,24,-9,-14,7,
-3,3,-9,0,2,-9,0,3,0,0,9,9,9,-9 Recipe ScoreBest Candidate C155 10,11,15,22,23,29,38,39,40,41,42 67
Worst Candidate C498 1,8,11,13,17,38,39,43 -50
Rule Knowledge MethodThe Method1. Randomly generate 5 – 10 candidate solutions
a) Sum the number of appearances of each rule in candSet and store in Rule Knowledge Matrix2. User evaluates candidates on a 1-5 scale based on likability (similar to DPM)
a) Project candidate scores onto rules and compute an average score for each rule
3. LEARN & GUIDE - Compute probabilities for each rule in Rule Knowledge Matrix and generate a new set of candidates based on these probabilities
4. Repeat 1 - 3
123456789
10111213...
n=154
Srfr
1 2 3 4 5 6 7 8 9 10 11 12 13 . . . n=154
23
8.623
LEARN
Rule Knowledge Method
Exploit – target rules that have consistently been rated highly in past
Explore – target rules with a low frequency value, f
B – tuner to control explore vs. exploit
Calculating probabilities
Probabilities
Rule 5
Rule 6
Rule 7
Pr
Rule 5Rule 6Rule 7
Questions
?