1. The Creative Value of Bad Ideas
A Computational Model of Creative Ideation
Ricardo Sosa
John S. Gero
2. Key ideas
1. Ideational fluency vs. quality variance
2. Simple computational models to support
insights for future research and practice
3. Assumptions about ideation:
1. Some ideas are easier to access than others
2. Ideas are connected to each other
4. Exploration and exploitation strategies
5. Distinguish ideas according to their potential to
generate new ideas
3. Idea accessibility
Likelihood to generate a particular idea or set of
ideas for a given design task: conformity metrics
(commonplace–original)
Idea connectivity
Likelihood of one idea leading to other ideas
(fitness landscapes, network linkages, trains of
thought)
4. Computational models of creativity
1. Generative models:
“Can computers ever be creative?”
2. Systemic models:
“Can computation help us understand creativity?”
5. Computational models of creativity
Contributions can inform other research
traditions by:
• Framing new hypotheses
• Testing the consistency and implications of
assumptions
• Proposing new experimental settings
In-vivo, in-vitro, in-silico studies of creativity
6. Computational model of ideation
“Use g geometries of s sides as the initial elements to
generate as many different compositions (g’s’) of more
than g geometries of equal or more s sides”
7. Computational model of ideation
{4,0,4,2,0, (3,3,3,4)} {4,1,3,0,2,4, (3,3,4,4)} {3,1,4,1,0, (3,3,6)} {3,2,3,1,0, (4,5,5)} {5,1,5,0,0, (3,3,3,5,5)}
Accessibility: given
by mean frequency
of solutions
(exploration) over
repeated cases
Exploitation: a
guided search
informed by rules
(‘design concepts’)
inferred from
experience
Connectivity:
mapping of design
concepts leading to
new solutions
(exploitation)
8. Findings
1. Choice of representation shown to influence
the type and degree of creativity
2. Exploration/exploitation rate of .25/.75 =
highest number of total ideas… due to higher
idea variances upon which exploitation builds
on novel ideas
9. 3. Ideation types:
– Solutions of low-accessibility leading to
solutions of high-accessibility: a “difficult
way to reach easy ideas”
– Low-accessibility solutions that lead to
new solutions of low accessibility too:
“uncommon ideas that yield other
uncommon ideas”
– High accessibility ideas connected to low
accessibility ideas: “easy shortcuts to reach
rare ideas”*
10. 4. Bad ideas can be very valuable for
creative ideation:
– Low-score or invalid solutions that lead to
valuable and uncommon solutions
Task: “to find 3 or more final geometries and shapes of 5 sides”
a) solution {2,1,3,2,0 (3,6)} and b) solution {4,1,3,2,0 (3,3,4,5)}
11. “The best way to get good ideas
is to get lots of ideas”
“…and the best way to get lots of
ideas is to first generate a few
that are as different as possible
and then strategically build on
them”
“…including some seemingly
bad ideas”
Linus Pauling, the only person
to be awarded two unshared
Nobel Prizes
A very simple
computational model
12. Discussion
Needed: research in facilitation techniques that
monitor idea accessibility –or variance metrics
Accessibility-based tools to balance
exploration/exploitation ratio
Connectivity-based tools to inform the
exploitation strategies
13. Discussion
New studies and metrics of ideation to
target the value of new ideas based on
their potential to trigger more ideas
Editor's Notes
Ideational fluency or quality variance? maximise the total number of ideas vs. the quality of one or a very small number of great ideasHelps to clarify and question assumptions, reveal insights, test hypotheses and in general help our thinking about creativity and potentially inspire innovative research and practice
Here, ideas of low accessibility that still satisfy the task’s goal can be considered as preferred over ideas that are commonly found across simulation ru. When r = 1.0, the idea space has a uniform accessibility distribution, whilst as r approaches 0, variance increases indicating that a few ‘gems’ are found in a ‘dense haystack’. ns
2. Combinatorial or derivative processes in ideation are more efficient when they act upon a highly diverse population of initial ideas