A Study on Multimemetic Estimation of Distribution Algorithms
1. A Study on Multimemetic Estimation of
Distribution Algorithms
Rafael Nogueras & Carlos Cotta
Memes are patterns-based rewriting rules
[CàA]:
• C, A ∈Σr with Σ={0,1,#}, r∈Ν
• ‘#’ represents a wildcard symbol
Meme à Unit of imitation
Encoded in computational
representations (meme↔gene)
MMA
Focus on meme
manipulation &
propagation
Best Fitness Meme Diversity Meme Success Rate
UMDA
PBIL
MIMIC
COMIT
TRAP-128
HIFF-128
HXOR-128
SAT-128
Let G=00010011, and let a
meme be 01#à1#0:
PPSN 2014
Ljubljana, Slovenia
Memes
Genes
MEME
Self-adaptive Search
Engine
EDA Cycle
1. Pop ß Sample(p(x))
2. pop’ ß Select(pop)
3. p(x) ß Update(pop’)
EDA learns the joint probability
distribution p(x) using the most
promising individuals at each
generation.
Wilcoxon ranksum
Multimemetic EDAs with elitism are superior to MMAs.
Memetic search process is better when no Laplace correction is
used in meme modeling.
Investigate other representation of memes.
More complex probabilistic graphical models (Bayesian Networks).
Decoupled evolutionary model.
EDA
Non-Elitist Elitist Laplace Non-Laplace Laplace Non-Laplace
Three symbols for problem/EDA respectively indicating how the
algorithm compares with its (non-)elitist counterpart, with sMMA, and
with the algorithm with the highest median for the corresponding
problem (which is marked with a star « in this third position). A
white/black circle (™/˜) for a worse/better median.
sMMA
Focus on meme
modeling
MMEDA
LC