Interactive Evolutionary Computation : a survey of existing theory Yann SEMET Compiegne University of TEchnology

IEC ? Interactive Evolutionary Computation Allow the user to play a role 2 definitions : Narrow : user = fitness function Broad : user controls the search Not restricted to GAs

Interactive Evolutionary Computation

Allow the user to play a role

2 definitions :

Narrow : user = fitness function

Broad : user controls the search

Not restricted to GAs

Motivation Intermediate paradigm of AI Assist user’s creativity Key applications : Computer graphics Data Mining Design

Intermediate paradigm of AI

Assist user’s creativity

Key applications :

Computer graphics

Data Mining

Design

Overview Background Menagerie of techniques : Interface enhancing Working on convergence Mathematical formalism General design heuristics

Background

Menagerie of techniques :

Interface enhancing

Working on convergence

Mathematical formalism

General design heuristics

A typical system

History and state of the art Begins in the mid 80’s (Dawkins) Peak in the early 90’s Among other names : Sims, Parmee, Banzhaf and Graf A predominant Japanese school : Takagi et al. A strongly application oriented field

Begins in the mid 80’s (Dawkins)

Peak in the early 90’s

Among other names : Sims, Parmee, Banzhaf and Graf

A predominant Japanese school : Takagi et al.

A strongly application oriented field

Toward a theory of IEC Identifying issues : Limited resources Noise GUI design questions Meeting psychologists and designers A set of tricks

Identifying issues :

Limited resources

Noise

GUI design questions

Meeting psychologists and designers

A set of tricks

2 issues : fitness input & display Takagi, 96 User = fitness function : narrow def. Issue : what scale ? 4 methods tested : A : Continuous input, sequential display B : Continuous input, simultaneous display P1 : Discrete input P2 : Hybrid discrete/continuous

Takagi, 96

User = fitness function : narrow def.

Issue : what scale ?

4 methods tested :

A : Continuous input, sequential display

B : Continuous input, simultaneous display

P1 : Discrete input

P2 : Hybrid discrete/continuous

Results

Fitness prediction Ohsaki et Takagi, 98 Idea : present individuals in the right order How : predict fitness based on previous generation Several methods : NN, distance, etc.

Ohsaki et Takagi, 98

Idea : present individuals in the right order

How : predict fitness based on previous generation

Several methods : NN, distance, etc.

Illustration

Fitness prediction results Good prediction But bad psychological impact ! Users prefer the conventional way Why : users don’t use global fitness but weighted sub functions Also : uniformity hinders the detection process

Good prediction

But bad psychological impact !

Users prefer the conventional way

Why : users don’t use global fitness but weighted sub functions

Also : uniformity hinders the detection process

Clustering Parmee 96 ;Kim et Cho, 2001 Evaluate less, breed more K-means clustering Fitness calculation based on distance

Parmee 96 ;Kim et Cho, 2001

Evaluate less, breed more

K-means clustering

Fitness calculation based on distance

Results for clustering

Psychometrical space Sugimoto et Yoneyama, 2000 Parametric topology vs psychometric topology Suggested solution : map the 2 spaces Tool : Kruskal’s M-D-Scal Then : fitness assignment by fuzzy reasoning

Sugimoto et Yoneyama, 2000

Parametric topology vs psychometric topology

Suggested solution : map the 2 spaces

Tool : Kruskal’s M-D-Scal

Then : fitness assignment by fuzzy reasoning

Result : robustness

Knowledge embedding Idea : allow active intervention Accelerates convergence, decreases useless disruption A first step toward linkage learning ? Takagi, 2000 Gene freezing

Idea : allow active intervention

Accelerates convergence, decreases useless disruption

A first step toward linkage learning ?

Takagi, 2000

Gene freezing

Convincing Results

Visualized IEC Takagi, 2000 Idea : using the human 2D global perception ability Solution : nD -> 2D mapping User performs selection

Takagi, 2000

Idea : using the human 2D global perception ability

Solution : nD -> 2D mapping

User performs selection

Results

Accelerating convergence Ingu et Takagi, 99 Idea : fitting a single peak function Quadratic approximation Might include past individuals Data selection issue

Ingu et Takagi, 99

Idea : fitting a single peak function

Quadratic approximation

Might include past individuals

Data selection issue

Results for single peak fitting

Mealy Automata Rudolph, 97 The only attempt of mathematical modelization A more general framework than Markov Chains : includes inputs What it brings ? Not clear yet ! Clues : decomposition, convergence studies

Rudolph, 97

The only attempt of mathematical modelization

A more general framework than Markov Chains : includes inputs

What it brings ? Not clear yet !

Clues : decomposition, convergence studies

THE slide The way to GA success : Know what GAs are processing – BBs Ensure an adequate supply of raw BBs Ensure growth of superior BBs Ensure the mixing of BBs Ensure good decisions among BBs Solve problems with bounded BB difficulty

The way to GA success :

Know what GAs are processing – BBs

Ensure an adequate supply of raw BBs

Ensure growth of superior BBs

Ensure the mixing of BBs

Ensure good decisions among BBs

Solve problems with bounded BB difficulty

10 steps for a perfect Creative Evolutionary Design System Bentley et O’Reilly, GECCO 2001 Find a domain in which it makes sense to use a computer for “creativity enhancement” Find a good reason for using a creative system at all. Negotiate appropriately balanced control Find a good niche for IEC in the design process Make it generative and creative

Bentley et O’Reilly, GECCO 2001

Find a domain in which it makes sense to use a computer for “creativity enhancement”

Find a good reason for using a creative system at all.

Negotiate appropriately balanced control

Find a good niche for IEC in the design process

Make it generative and creative

10 steps for a perfect Creative Evolutionary Design System II Make it understandable Have an easy and effective way of evaluating solutions Find people who are actually prepared to use the system Get lots of money to pay R&D costs Start a company and make a billion.

Make it understandable

Have an easy and effective way of evaluating solutions

Find people who are actually prepared to use the system

Get lots of money to pay R&D costs

Start a company and make a billion.

Summary 2 definitions of IEC 1 main issue : reduce human fatigue 7 tricks 1 mathematical attempt General heuristics

2 definitions of IEC

1 main issue : reduce human fatigue

7 tricks

1 mathematical attempt

General heuristics

Conclusions & discussion A young field, strongly application oriented Needs : More formalization More understanding : formalize issues such as linkage, hybridization, etc. Computational creativity & innovation ?

A young field, strongly application oriented

Needs :

More formalization

More understanding : formalize issues such as linkage, hybridization, etc.

Computational creativity & innovation ?