Architecture for Exploring Large Design Spaces

1
Architecture forExploring Large Design Spaces

• John R. Josephson, B. Chandrasekaran,
• Mark Carroll, Naresh Iyer,
• Bryon Wasacz, Qingyuan Li,
• Giorgio Rizzoni, David Erb

2
Architecture for exploring large design spaces
Three synergistic components
Seeker
Filter
Viewer
3
Design Seeker

• Human initiates automated design search which may
  work by considering combinations of
• generic devices (configurations)
• alternative components
• representative parameter values.
• Designs are evaluated according to multiple
  criteria using simulation-based and other critics

4
Design Seeker
Device Library
Critics
Search control
Constraints
Evaluated designs
5
Big search !

• Search may be massive and exhaustive.
• Largest experiment to date
• 2,152,698 designs were generated and evaluated,
  of which 1,796,025 were fully specified.
• Each fully specified design was evaluated using
  multiple simulations.
• Seeker used idle time on 209 workstations to
  search the space in 6.8 days (wall-clock time).
  (The maximum number running at any one time was
  159.)

6
Dominance Filter
Dominance algorithm
7
Dominance Filter

• Design candidate A is said to dominate candidate
  B if A is superior or equal to B in every
  criterion of evaluation and strictly superior for
  at least one criterion.
• Dominated designs are removed. (This is lossless)
• Surviving designs are Pareto optimal (improvement
  on any criterion will reduce value on another)
• Tolerances may be specified for the comparisons.

8
Dominance Filter
Dominance algorithm
Dominance filtering can be very effective.
9
Effectiveness of dominance filtering
Using 4 criteria and reasonably realistic
simulation models
Dominance filtering is very effective! Dominance
filtering scales very well!
10
Efficiency of dominance filtering algorithm
1,796,025 1,078
4.5 hours (serial post processing)
11
Effect of number of criteria
In experiment B with 17,711 designs
The effectiveness of dominance filtering
apparently tends to decrease as the number of
criteria increases.
12
Interactive Viewer
Filter
Viewer
Tradeoffs are explored interactively.
13
Interactive Viewer

• visualization of trade-offs
• zooming to selected regions in trade-off space
• selection of subsets by structural constraints
  (not implemented)
• initiation of more focused search (not
  implemented)
• initiation of additional search, e.g., add
  criteria (not implemented)

14
Visualizing search results
15
Visualizing search results
16
Visualizing search results
17
Visualizing search results
18
Visualizing search results
19
Visualizing search results
20
Visualizing search results
21
Visualizing search results
22
Visualizing search results
23
Exploring large design spaces
Human-in-the-loop multi-criterial optimization
Seeker
Filter
Viewer
24
Patent application has been submitted.
25
Next Steps

• Technology for composable simulation models
• Improved viewer - more types of displays
• Automatic extraction of generalizations

26
Questions?
27
Design Seeker

• Essentially
• a generator of design
• evaluators for designs

28
More generally

• The Seeker consists of
• a generator of choice alternatives
• evaluators for choice alternatives

29
Seeker based on client-server