Abstract.
In order to improve the behavior of Particle Swarm
Optimization (PSO), the classical method is often extended by additional
operations. Here, we are interested in how much ``PSO" remains in this
case, and how often the extension takes over the computation. We study
the variant of PSO that applies random velocities (then called forced
moves) as soon as the so-called potential of the swarm falls below a
certain bound. We show experimentally that the number of iterations the
swarm actually deviates from the classical PSO behavior is small as long
as the particles are sufficiently far away from any local optimum. As soon
as the swarm comes close to a local optimum, the number of forced moves
increases significantly and approaches a value that depends on the swarm
size and the problem dimension, but not on the actual fitness function,
an observation that can be used as a stopping criterion. Additionally,
we provide an explanation for the observed phenomenon in terms of the
swarms potential.