Lecture 6 Flashcards
Compared to EDA, what is the problem with variation in simple GA?
Thier are blind to the problem structure. Which leads to inefficiency of the variation.
What does it mean that a function is additively decomposable?
When the problem can be decomposed into smaller problems, of which the sum of their solutions equals the solution of the larger problem.
f.i. functions whos fitness depends on blocks
What are block identifiers? (and how are they represented)
The parameters in which blocks are defined.
k is block length
m is # blocks
l = k*m is string length
Explain what thight encoding means.
It is when the genes in each block are neighbours:
0000 1111 2222 …
Explain what loose encoding means.
It is when there is a single gene of each block present when you take a range equal of the block lenght:
012345 012345 …
What does disruption mean?
It means that blocks get broken during variation.
When is a function deceptive?
(from slides)
Any schema of size smaller than the block size of the
deceptive trap function has a better average fitness
if it has more zeros.
Following the schema theorem, those “0-schemata”
have better chance at receiving an increasing
number of matches over time.
This is only false for full-order schemata.
In Empirical scalability analysis, what is the definition of succes?
When the optimal solution is found in a certain percentage of runs.
What do we try to solve using Empirical scalibilty analysis?
Given a value for ℓ, what is the minimally required population size for wich the EA solves the problem?
Decribe how empirical scalibiltity analysis works and what criterea needs to be met.
Try to increase population size (exponentially) until success has been reached. This will be the upper bound. Then tak the lower bound as half of your upper bound and perform additionaly search within this range in order to find the best population size.
This process should be repeated a couple times to verify results
What is the principle of Model based EAs?
Appart from being an EA, it tires to learn and exploit structures in the problem.
Give a rough psudo code of how model based EAs operate
- The look at the last generation
- They learn a stochastic model based on that
- They generate new solutions that allign with that model
- Perform survivor selection.
- repeat
What does EDA stand for?
Estimation-of-Distribution Algorithms
Why is EDA a model based EA?
Because it uses the previous generation to estimate the distribution of the next generation.
What major benefit does EDA have?
The remaining search space shrinks very fast, so very few generations needed to converge.
Explain why we would want to used factorized EDAs
The hypothesis is that many problems can be decomposed in sets of vairables that togather have an above average contribution to a solutions fitness.
When we’re learning the distribution estimation in EDAs it can be much more efficient when we apply factorization on the distribution to capture this
What is univariate factorization in EDA?
(aka univariate EDA)
the distribution over each variable is modelled to be
independent from that over every other variable
Explain how UMDA operates.
UMDA is a type of EDA:
It evaluates a larger subselection of the generation and tries to estimate the chances of each binary gene to be present in the optimal solution based on the fitness of the genotypes. Then the next generation is sampled from those distributions.
Does UMDA concidered decomposable?
Yes, its distribution uses univariate factorization as each variable is independently evaluated.
Are univariate EDAs the solution for deceptive traps?
No, the univariate EDA will favor the solutions with above average fitness and converge very fast towards the trap.
Unless the population size is very large and it can therefor generate many optimal solutions at initialisation.
