Slide 1 - Boris Burdiliak
Download
Report
Transcript Slide 1 - Boris Burdiliak
Evolutionary Design (2)
Boris Burdiliak
Topics
Representation
Multiple objectives
Representation – Introduction
EA are based on natural evolution
evolution provides a quick and easy way
to solve difficult problems (scheduling,
machine learning, ordering problems, data
mining, control, design optimisation)
Representation – Problems
traditional implementations of evolutionary
search suffer from the same drawbacks as
all conventional search algorithms
they rely on a good parameterization to
find a good solution
evolution is limited by the representation it
can modify
Explorative evolution
evolution can search for good search
spaces, even as it searches within a space
dimensionality of a space, the
parameterization, the representation of
solutions
solutions = inventions, not improvements
Exploring the Explorer
what is evolution doing when it
explores/optimises?
fixed-length parameterization
evolution = optimisation
parameters define a set of components
explores new ways of constructing a solution
A Framework for Expl. Evolution
EA
GA/GP (genotype/phenotype distinction)
EA/ES (no distinction)
prob.: multiple objectives, premature
convergence, changing fitness functions
Genotype Representation
defines the search space(defines components)
prob.: dissimilar problems are close to e.o.,
disruption of inheritance
A Framework for Expl. Evolution 2
Embryogeny
Phenotype Representations
mapping process from genotype to phenotype
provide the mechanism for constructing whole solutions from
components
external: programmer writes the software that performs the
mapping
explicit: every step of growth process explicitly held as a part of
genotype, and evolved
implicit: set of rules (encoded as bin. strings in genotype)
allow direct evaluation of fitness function
Fitness functions
provide evaluation score for every solution
Multi objectives
optimise function F(x) under additional
constraints, i.e.
max {x} F(x)
g_1(x,p) <= 0, … g_l(x,p) <= 0
p(p_1,…,p_u) – real-valued parameters
relative importance, interactions of
objectives (also contradictory)
Pareto optimisation
Def: We will say that a point xD Paretodominates a point yD with respect to function
F, denoted yx, if {i=1..k} (f_i(y) f_i(x)) and
at least one of inequalities is strict
We say that point x_pD is Pareto-optimal or
non-dominated (for a given function F) if there is
no point yD that Pareto-dominates x.
Set FD is called Pareto front with respect to
function F if every element x F is Pareto optimal
with respect to function F
Selection phase
not all elements from the Pareto front are of a
good fitness value
Pareto tournament
Pareto sort
standard tournament selection (best individual
according to Pareto ordering)
sort according to Pareto ordering
if both ind. are dominant, sort acc. to fitness value
Pareto truncation
choose parent only among the non-dominated
#of non-dominated elements
population size = 100
standard tournament (size 2)
averaged over 50 runs
optimising 3 parameters
non-dominated elements
after
after
after
after
50 runs: cca 42%
150 runs: >50%
350 runs: almost 70%
500 runs: cca 85%
Pareto front is too large
Pareto ranking
Def: Pareto rank r in a set X={x_1,…,x_n}
is assigned in a following way
(xX) r(x) 1
("x{x_1,…x_n}) ("y{x_1,…,x_n} \ {x})
if (r(x)=r(y) x > y) r(x) r(x)+1
if (r(x)=r(y) x < y) r(y) r(y)+1
Conclusions
requirement is for methods that can work
with a qualitative in addition to a
quantitative characterization of importance
preference methods
fuzzy set methods
agent based systems which model humanbased solution evaluation process
Fuzzy preferences
transforming qualitative relationships between
objectives in multi-objectives optimization into
quantitative attributes
characterisations for every 2 objectives:
much less important <<
less important <
equally important ==
don’t care #
more important >
much more important >>
Don’t care relation
don’t care is treated exactly as equally
important
hesitations
weak preference
can not decide: a > b and a ==b, being sure that
not (b > a)
incomparability
can not decide: a > b and b > a
Method used
k objectives – k(k-1)/2 questions -> construct
the fuzzy preference relation R
compute complete oredering on A using the
relation graph G=(A,R)
define relations:
== equally important
< is less important
<< is much less important
~ not important
! important
Algorithm
set of objectives O
construct the equivalence classes C_i
(relation ==), choose one element x_i
from each class X={x_1,…,x_m}
…
for each xi compute weight w(xi)
Weighted Pareto method
Definition of non-dominated vector
Definition of w-non-dominanace
Definition of (w, tau)-non-dominance
THE END