Computer Science Evolutionary Computation Genetic Algorithms Memetic Algorithms Island-based Algorithms Complex Networks Complex Systems

1. Evaluating Island-based EAs on Unstable Networks
with Complex Failure Patterns
Rafael Nogueras & Carlos Cotta
Dpto. Lenguajes y Ciencias de la Computación, Universidad de Málaga
GECCO 2017
Berlin
EphemeCH
TIN2014-56494-C4-1-P
Use of parallel and distributed models of EAs to improve solution quality and reduce computational times.
The island model spatially organizes populations into partially isolated panmictic demes.
Two emergent computational environments: P2P networks and desktop grids.They are dynamic and unstable.
Churn: the combined effect of multiple computing nodes leaving and entering the system along time.
Scale-free topology (Barabási-Albert model)
Platforms with many nodes
Instability
FaultTolerance
• Unstable scenarios with correlated
failures pose a hard challenge.
• Self-✭ properties are an effective
solution.
• Self-scaling and self-sampling are
robust under different failure
patterns.
FutureWork
Extend this work on other models of
correlated failures and network
topologies.
Self-Scaling: the algorithm changes its structure in response to variations in the
problem or environment. Islands dynamically change their size in the presence of churn:
• If a neighboring island goes down, an island increases its size.
• Active islands exchange individuals in order to balance their sizes (self-balancing).
Self-Sampling: Prevent randomness when an island increases its size ⇒ EDAs to
estimate the population of each island to be enlarged generating new individuals by
sampling.
Node failures are influenced by neighboring
nodes ⇒ sandpile model variant.
A micro-failure event happens with probability
p(t): when the number of such micro-failures
equals the number of active neighbors ⇒ the
node is disconnected from the system.
Reactivation doesn’t change.
Each node can switch from active to inactive or
vice versa independently of other nodes with
some probability p(t) that depends on the time
it has been in its current state and follows a
Weibull distribution.
TRAP HIFF MMDP
Scale-free networks feature
a power-law distribution in
node degrees.These networks
are often the result of
processes driven by
preferential attachment.