/**
* Just reevaluate the {@link Entity entities} inside the topology. An entity's implementation
* should handle updating anything else that is necessary, e.g. a {@link StandardParticle}'s
* <code>personal best position</code> in the case of {@link PSO}.
*
* <p>{@inheritDoc}
*/
@Override
public void performReaction(E algorithm) {
Topology<? extends Entity> entities = algorithm.getTopology();
int reevaluateCount = (int) Math.floor(reevaluationRatio * entities.size());
reevaluate(entities, reevaluateCount);
}
/**
* Respond to environment change by re-evaluating each particle's position, personal best and
* neighbourhood best, and reinitializing the positions of a specified percentage of particles.
*
* @param algorithm PSO algorithm that has to respond to environment change
*/
@Override
public void respond(E algorithm) {
Topology<? extends Entity> topology = algorithm.getTopology();
// Reevaluate current position. Update personal best (done by reevaluate()).
Iterator<? extends Entity> iterator = topology.iterator();
int reinitCounter = 0;
int keepCounter = 0;
int populationSize = algorithm.getTopology().size();
while (iterator.hasNext()) {
DynamicParticle current = (DynamicParticle) iterator.next();
ZeroTransformation zt = new ZeroTransformation();
// makes sure the charged particles are randomly positionned accross the topology
if (reinitCounter < Math.floor(populationSize * reinitialisationRatio)
&& randomiser.nextDouble() < reinitialisationRatio
&& current != (algorithm).getTopology().getBestEntity()) {
current.getPosition().randomize(this.randomiser);
current
.getProperties()
.put(EntityType.Particle.VELOCITY, Vectors.transform(current.getVelocity(), zt));
current
.getProperties()
.put(EntityType.Particle.BEST_POSITION, current.getPosition().getClone());
++reinitCounter;
} // if
else if (keepCounter > Math.floor(populationSize * (1.0 - reinitialisationRatio))
&& current != (algorithm).getTopology().getBestEntity()) {
current.getPosition().randomize(this.randomiser);
current
.getProperties()
.put(EntityType.Particle.VELOCITY, Vectors.transform(current.getVelocity(), zt));
current
.getProperties()
.put(EntityType.Particle.BEST_POSITION, current.getPosition().getClone());
++reinitCounter;
} // else if
else {
++keepCounter;
} // else
}
// Re-evaluate:
reevaluateParticles(algorithm); // super class method
}
/**
* Reinitialises all the output weights of a cascade network within a PSO.
*
* <p>{@inheritDoc}
*/
@Override
public void performReaction(E algorithm) {
NNDataTrainingProblem problem = (NNDataTrainingProblem) algorithm.getOptimisationProblem();
NeuralNetwork network = problem.getNeuralNetwork();
int precedingLayersSize =
network.getArchitecture().getArchitectureBuilder().getLayerConfigurations().get(0).getSize()
+ network
.getArchitecture()
.getArchitectureBuilder()
.getLayerConfigurations()
.get(1)
.getSize();
if (network.getArchitecture().getArchitectureBuilder().getLayerConfigurations().get(0).isBias())
precedingLayersSize++;
int outputLayerSize =
network
.getArchitecture()
.getArchitectureBuilder()
.getLayerConfigurations()
.get(2)
.getSize();
int nrOfweightsToDo = precedingLayersSize * outputLayerSize;
Topology<? extends Entity> entities = algorithm.getTopology();
for (Entity entity : entities) {
DynamicParticle particle = (DynamicParticle) entity;
Vector position = (Vector) particle.getPosition();
Vector velocity = (Vector) particle.getVelocity();
for (int curElement = position.size() - nrOfweightsToDo;
curElement < position.size();
++curElement) {
((Real) position.get(curElement)).randomize(randomGenerator);
velocity.setReal(curElement, 0.0);
}
}
}
